acceptodds
Under review as a conference paper at ICLR 2027

HieraNTP: Structured Risk Representations for Time-Series Anomaly Detection via Hierarchical Next-Token Prediction

Abstract

Time-series anomaly detection commonly converts deviations from learned temporal regularities into anomaly scores. This practice can conflate violations of temporal regularity with task-specific anomaly judgments: the same deviation may be anomalous under one task definition and benign under another. Aggregating heterogeneous deviations into a single score before task-specific adaptation can obscure the evidence needed to distinguish these cases. We introduce HieraNTP, a Hierarchical Next-Token Prediction framework that separates self-supervised predictive modeling from task-specific risk interpretation. HieraNTP pretrains a hierarchical tokenizer and a causal next-token predictor to capture temporal regularities at multiple scales. It then organizes prediction discrepancies into a structured risk representation, preserving complementary evidence across temporal scales in token, local-statistic, and observation spaces. A task-adaptive detector uses anomaly labels to select and combine this evidence for the target task. Across anomaly detection benchmarks, HieraNTP improves VUS-PR by 12.0% on TSB-AD-U and 15.5% on TSB-AD-M relative to the strongest baselines, and achieves the highest composite score on both TAB tracks, with controlled comparisons supporting the value of retaining distinct evidence.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Anomaly Detection, Time Series Representation, Task-Adaptive Detection, Structured Risk Representations

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SaMuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper identifies the "predictive risk–decision gap" in time series anomaly detection: scoring directly by prediction/reconstruction error may cause the same deviation to be anomalous under one task definition and normal under another. To address this, the paper proposes HieraNTP: it first self-supervisedly pretrains a "tokenizer + next-token predictor," then freezes them, and uses a labeled adaptive detector (XGBoost, comprising a shared base detector plus dataset-specific extensions) to learn evidence combination. On TSB-AD-U/M, VUS-PR improves by 12.0%/15.5% respectively over the strongest baseline, and it achieves the highest overall score on both tracks of TAB.

Critical strengths
  1. The problem framing is insightful: the discussion of the "predictive risk–decision gap" is clear and well-supported
  2. The experiments are extremely solid: dual benchmarks and experiment designs. And the results are leading.
Critical weaknesses
  1. Imbalance between complexity and interpretability: the thousands-of-dimensions handcrafted evidence representation + XGBoost make the engineering implementation complex。
  2. Strong dependence on labels: the task-adaptive detector requires task labels, and when comparing with unsupervised/zero-shot baselines, the differences in setup need to be clarified.
  3. Generalization is not sufficiently validated: pretraining uses QUALS and evaluation uses TSB-AD/TAB, with no independent validation on completely unseen/cross-domain datasets; on some datasets (TODS, SMD), it still underperforms TimeRCD.
Questions for authors
  1. What are the computational/storage/inference overheads of the thousands-of-dimensions evidence representation? Could you provide a runtime and GPU memory comparison, or validate the performance after feature dimensionality reduction (e.g., top-k features, PCA)?
  2. The task-adaptive detector relies on labels. Is the comparison with unsupervised/zero-shot baselines (TimeRCD, MOMENT, etc.) fair?
  3. What are the trade-offs between XGBoost and an end-to-end differentiable detector? Have you considered replacing it with a lightweight MLP/graph model?
  4. On completely cross-domain distribution data, how does the framework perform?
8 comments

To W1Q1: Representation Dimensionality and the Computational, Storage, and Inference Costs of Evidence Processing

Thank you for raising the concerns about representation complexity and computational overhead. We added top-k feature selection, PCA experiments, and runtime measurements.

HieraNTP expands 64 base discrepancy channels through three processing paths and 13 temporal summaries, yielding 2496 univariate features. Retaining variable means and maxima gives 4992 multivariate features. Each feature has an identifiable evidence source and processing configuration.

Top-k performs post-hoc selection using split gain from the Tuning-trained detector. PCA is fitted on training data, followed by detector retraining on the projected features. The table reports all completed settings using mean VUS-PR over the complete Eval sets.

Retained dimensions KU: Top-kU: PCAM: Top-kM: PCA
Full representation65.56—50.26—
6464.1753.7847.0444.76
12864.5254.1348.3346.24
25665.2755.1148.8346.30
51265.1455.3649.6146.51
102465.7055.6249.7946.55

VUS-PR is reported in percent. With 1024 retained features, top-k achieves 65.70% on U and 49.79% on M, showing that the supervised detector concentrates its decisions on a small subset of task-relevant evidence. Retraining with 1024 PCA components achieves only 55.62% and 46.55%, indicating that a common PCA projection loses task-relevant risk information distributed across channels. Different tasks require different evidence, motivating HieraNTP to preserve these channels for supervised selection and combination.

For computation and inference, complete U350 and M180 evaluations take approximately 2.8 and 20.4 minutes using seven RTX 4090 GPUs and 48 workers. These evaluations cover 530 test sequences and approximately 264 million scalar observations. Single-sequence inference uses approximately 890 MiB of GPU memory. Under the concurrent configuration, peak aggregate memory usage per GPU is approximately 20.74 GiB.

For storage, the XGBoost model files occupy approximately 11.13 MiB for U and 8.50 MiB for M. Intermediate evidence is passed through memory, and the final FP32 anomaly-score arrays occupy approximately 69.28 MiB for U350 and 74.73 MiB for M180. We describe the representation dimensions and reference feature selection in Section 3.3, report the top-k and PCA comparisons in Appendix E.6 (Table 33), and document runtime, GPU memory, and storage costs in Appendix C.3.

To W2Q2: Fairness of Comparisons Between Supervised Task Adaptation and Unsupervised or Zero-Shot Anomaly Detection

Thank you for raising the question about supervision and comparison fairness. The TSB-AD Tuning sets and TAB training data provide labels that support supervised anomaly detection. HieraNTP follows these training and test splits, uses training labels to fit the task-adaptive detector, and evaluates on the official test sets. The manuscript also identifies the zero-shot and task-adapted settings of the methods.

Our use of supervision is motivated by the predictive risk–decision gap. Predictive discrepancies identify violations of temporal regularity, while anomaly labels specify which violations require detection. The Yahoo case illustrates this distinction. Missing peaks and disrupted cycles both produce risk responses, but the task labels only the latter as anomalous.

To compare performance under matched supervision, we organized the available predictive discrepancies, reconstruction discrepancies, and representations from TSPulse, Chronos-Bolt, and MOMENT, then applied the same anomaly labels and XGBoost training pipeline as HieraNTP. The table reports VUS-PR on TSB-AD-U/M.

MethodZero-shot detectionTask-adaptive detection
TSPulse0.48 / 0.360.57 / 0.49
MOMENT0.38 / —0.52 / 0.49
Chronos-Bolt0.29 / —0.57 / 0.49
Multiscale statistics—0.46 / 0.41
HieraNTP—0.66 / 0.50

With task adaptation, the three pretrained models reach 0.52–0.57 on U and approximately 0.49 on M, competitive with leading detectors in the main comparison. These results show that retaining multiview risk evidence and learning task criteria from training labels benefits different pretrained models. Under matched supervision, HieraNTP achieves the highest VUS-PR among the evaluated representations on both tracks. This matched-supervision comparison has been added to Table 4 of the revision.

Our motivation is that forecasting accuracy and anomaly-detection accuracy differ in their task semantics. Anomaly annotations express the events that matter to the application. HieraNTP learns which risks and combinations should trigger anomaly decisions, connecting structured predictive evidence to the target task’s criteria.

The revision clarifies the supervision protocol in Section 4.1 and reports the matched-supervision comparison in Section 4.3 (Table 4), with feature construction and complete results documented in Appendix E.1 (Tables 23–24).

To Q3: Trade-Offs Between Gradient-Boosted Decision Trees and End-to-End Differentiable Neural Network Detectors

Thank you for this suggestion. We replaced XGBoost with a lightweight multilayer perceptron (MLP) and a graph attention network (GAT), using the same frozen predictor, structured risk representation, and anomaly-label supervision. Both alternatives retain shared training and dataset-specific adaptation. The GAT uses risk-evidence channels as nodes to learn interactions among evidence.

DetectorTSB-AD-U VUS-PR (%)TSB-AD-M VUS-PR (%)
XGBoost65.5650.26
MLP53.9948.72
GAT54.6943.70

XGBoost achieves the best performance on both tracks. Its tree splits select evidence according to input conditions, allowing different samples to use different evidence combinations. This matches HieraNTP’s objective of learning task-specific criteria from multiview predictive risk.

Feature masking is applied during detector training. We subsequently evaluate evidence dependence on a held-out univariate probe through changes in the anomaly logit under feature-wise masking. On the U held-out probe, the median share of normalized logit sensitivity carried by the top 256 features is 96.36% for XGBoost, compared with 34.51% for MLP and 39.32% for GAT. Thus, the trained XGBoost detector concentrates its decisions on a small subset of evidence, while the trained neural alternatives remain sensitive to more channels. Together with detection performance, these results support XGBoost’s effectiveness in selecting and combining task-relevant evidence under the current supervision.

The trade-off is that XGBoost learns conditional evidence selection through a separately trained detector, whereas MLP and GAT provide differentiable architectures for joint optimization with upstream representations. The replacement experiments support our use of XGBoost, while jointly training the predictor and a differentiable detector remains a promising direction. We have reported the detector comparison and feature-masking sensitivity analysis in Appendix E.7 (Table 34), with the connection to the task-adaptive detector described in Section 3.4.

To W3Q4: Cross-Dataset Generalization and Performance on Unseen Source Distributions

Thank you for raising the concerns about cross-domain generalization and dataset-level differences. Cross-domain anomaly detection involves transferring temporal regularities and accommodating changes in anomaly criteria. The same deviation can be normal or anomalous under different task definitions, as emphasized by the predictive risk–decision gap. HieraNTP learns reusable temporal knowledge through pretraining and uses downstream supervision to learn what the resulting risks mean for the target task.

Pretraining uses QUALS, derived from GIFT-Eval Pretrain, with no data leakage from TSB-AD or TAB. It trains the tokenizer and predictor, while downstream training labels supervise anomaly decisions.

To evaluate cross-dataset performance, we added leave-one-source-dataset-out evaluation. Each target source is excluded entirely from detector training and hyperparameter selection. The shared base detector is trained using the remaining sources and evaluated on the held-out dataset. The table reports mean VUS-PR over the complete TSB-AD Eval sets.

MethodTraining settingTSB-AD-U (%)TSB-AD-M (%)
HieraNTPStandard task adaptation65.5650.28
HieraNTPShared base detector only61.4849.91
HieraNTPLeave-one-source-dataset-out53.7036.89
TimeRCDZero-shot52.2431.68
TSPulseZero-shot48.0136.10

With the target source excluded, HieraNTP achieves 53.70%/36.89%, exceeding the zero-shot TimeRCD and TSPulse results. This shows that evidence–label relationships learned from other datasets support detection on held-out sources. The further gains from target-source training data demonstrate the value of task supervision in interpreting predictive risk.

Under standard evaluation, HieraNTP improves VUS-PR by 12.0% on TSB-AD-U and 15.5% on TSB-AD-M relative to the strongest baselines. TimeRCD remains stronger on TODS and SMD, showing that gains vary across datasets and suggesting differences in how learned anomaly criteria transfer from training data to test sequences. We have added the held-out source results to Section 4.3 (Table 3) and detailed the LOSO protocol and per-source results in Appendix E.3 (Tables 25–27), while retaining the dataset-level comparisons in Section 4.2 (Table 1).

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

HieraNTP separates self-supervised temporal modeling from task-specific anomaly interpretation. A pretrained multiscale tokenizer and hierarchical next-token predictor generate discrepancies in token spaces. These discrepancies are organized through reference calibration, temporal summaries, and variable aggregation, then supplied to a supervised XGBoost detector. The paper reports strong results on TSB-AD and TAB benchmark.

Critical strengths
  • The distinction between predictive deviations and task-defined anomalies is well motivated and practically relevant.
  • The framework coherently integrates complementary predictive evidence, and the empirical results support its effectiveness for anomaly detection.
Critical weaknesses

The contribution of dataset-specific supervision remains insufficiently characterized. The task-adaptive detector (XGBoost) uses TSB-AD Tuning labels from source datasets also represented in Eval, leaving unclear how much the reported gains depend on learning dataset-specific anomaly characteristics. A leave-one-source-dataset-out evaluation, excluding the held-out dataset from detector training and hyperparameter selection, would better establish generalization to unseen datasets.

Questions for authors
  • Could you report leave-one-source-dataset-out results, with task-adaptive detector training performed exclusively on the remaining datasets?
  • How would the framework accommodate benign concept drift, such as persistent changes in level, variance, or periodicity that become part of normal operating behavior?
  • Could you analyze multivariate false positives and false negatives and explain how evidence construction and variable pooling contribute to these errors? A controlled example in which univariate marginal distributions remain unchanged while cross-variable dependencies shift would help clarify the framework’s ability to detect relational anomalies.
5 comments

To W1Q1: Generalization to Unseen Source Datasets Excluded from Downstream Detector Training and Hyperparameter Selection

Thank you for suggesting this evaluation. We conducted leave-one-source-dataset-out (LOSO) experiments on both TSB-AD tracks to assess detection on datasets excluded from downstream training and hyperparameter selection.

Each fold holds out an entire source dataset, such as YAHOO or SMAP. With the pretrained modules frozen, we refit the shared token-space precision matrices and train the task-adaptive detector using only the remaining Tuning datasets. Hyperparameters are selected through three-fold validation grouped by sequence or entity within these retained datasets. The selected configuration then fits a 384-tree shared base detector, which scores the held-out dataset. Inference follows the paper’s offline procedure, including unlabeled reference-prefix calibration and bidirectional scoring.

We completed 19 held-out folds on TSB-AD-U and 11 on TSB-AD-M, covering 346 and 173 Eval sequences, respectively. Their sequence-mean VUS-PR scores are 53.93% and 37.47%. For datasets already absent from Tuning, we use a shared base detector trained on all available Tuning datasets. Combining these predictions covers the complete 350 univariate and 180 multivariate Eval sequences.

Table R2. Sequence-mean VUS-PR (%) on the complete TSB-AD Eval sets.

MethodSettingTSB-AD-UTSB-AD-M
HieraNTPStandard setting65.5650.28
HieraNTPShared base detector only61.4849.91
HieraNTPLeave-one-source-dataset-out53.7036.89
TimeRCDZero-shot52.2431.68
TSPulseZero-shot48.0136.10
TSPulseFine-tuned55.1638.66
TimeRCD+MAFTFine-tuned58.56—

Under LOSO, HieraNTP exceeds the zero-shot TimeRCD and TSPulse baselines on both tracks. Its scores remain below fine-tuned TSPulse on both tracks and TimeRCD+MAFT on U. These results show that evidence–label relationships learned from other datasets support detection on held-out datasets, while target-data adaptation provides further gains.

The comparison with Shared base detector only also characterizes where dataset-specific supervision contributes. That setting trains the shared detector across all Tuning datasets, including the target dataset when available. Its higher performance than LOSO indicates that target-dataset training data improves the shared decision rules, in addition to the gains from dataset-specific extensions. This finding supports HieraNTP’s separation of predictive modeling and task-specific risk interpretation. The predictive representation supplies common evidence, and downstream supervision learns how that evidence relates to the anomaly criterion.

We have incorporated LOSO into Section 4.3 (Table 3) and documented the held-out protocol, validation procedure, complete metrics, and per-dataset results in Appendix E.3 (Tables 25–27).

To Q2: Detection of Persistent Benign Changes in Normal Operating Regimes or Benign Concept Drift

Thank you for raising this question. We examine three mutually exclusive annotation patterns for persistent regime changes, covering normal behavior throughout, anomalies confined to the transition, and anomalies spanning the transition and subsequent state. The first two accept the subsequent regime as benign, with the second additionally requiring detection of the transition. The third requires continued detection of the changed state, including its predictable portions. These task-dependent judgments are encoded in human annotations and define what constitutes benign concept drift.

Such annotation differences coexist in TSB-AD. Figure 6 in the revised manuscript examines GHL 039 from TSB-AD-M and SMAP 545 and Exathlon 830 from TSB-AD-U. Post-hoc analysis of GHL 039 reveals predictive risk responses to the level shift. GHL's dataset-specific extensions make no splits on the three displayed risk channels, and the complete detector correctly classifies the displayed interval as normal. The table lists one representative signal from each case.

CaseLabels, transition / subsequent stateRepresentative risk signalSplit gain in dataset-specific extensionsShared base detector → Task-Adaptive Anomaly Detection
GHL 039Normal / NormalPatch variance, Patch 256, Mean (257)0.00%Specificity 100.00% → 100.00%
SMAP 545Anomalous / NormalPatch variance, Patch 8, Max (17)47.59%VUS-PR 99.52% → 99.97%
Exathlon 830Anomalous / AnomalousPatch mean, Patch 8, Mean (129)20.01%VUS-PR 87.06% → 100.00%

Results use the displayed windows. Task-Adaptive Anomaly Detection in the final column uses the complete ensemble with dataset-specific extensions. Specificity uses a threshold of 0.5, and split gain is normalized within the dataset-specific extensions.

The shared base detector already suppresses alarms for the benign change in GHL 039. Dataset-specific extensions further improve VUS-PR by 0.45 percentage points on SMAP 545 and 12.94 points on Exathlon 830. In SMAP 545, the detector responds to the anomalous transition and lowers its scores in the subsequent normal state while time-domain errors remain high. In Exathlon 830, persistent patch mean evidence supports high anomaly scores as some other discrepancies weaken. Time-domain error scoring achieves 6.93% and 92.38% window VUS-PR on these two cases, respectively.

Uray et al. (2026) [1] similarly observe that TimesFM's errors can diminish within persistent anomalies as their observations enter the conditioning context, with stronger responses near the boundaries. Zhu et al. (2026) [2] demonstrate the value of maximum aggregation across short- and long-horizon errors. We interpret these observations as practical manifestations of the predictive risk–decision gap. Error magnitude and persistence depend on the conditioning context and forecast horizon, while task criteria determine which portions of a regime change require detection.

HieraNTP addresses the broader question of how predictive risk should inform task-specific anomaly decisions. It retains responses across patch lengths and temporal summaries in a structured risk representation at each timestamp. XGBoost learns tree splits that select evidence channels and their nonlinear interactions using training anomaly labels. This enables task-adaptive detection to suppress alarms for benign changes, identify anomalous transitions, or sustain responses to persistent anomalies according to the target criterion. We have incorporated the regime-change comparison into Section 4.5 (Figure 6) and documented the signal definitions, shared-versus-adapted results, and additional visualizations in Appendix F.5 (Figure 12 and Table 36).

[1] Uray et al. Exploring Zero-Shot Foundation Models for Multivariate Time Series Anomaly Detection. EUROCAST 2026

[2] Zhu et al. When Foundation Models are One-Liners: Limitations and Future Directions for Time Series Anomaly Detection. [ICLR 2026]

To Q3: Multivariate False Alarms and Changes in Cross-Variable Dependencies

Thank you for this question. First, multivariate detection on TSB-AD and TAB introduces cross-variable evidence aggregation into HieraNTP. We independently construct multiscale predictive evidence for each variable, then pool it using variable means and maxima to form shared evidence channels for the task-adaptive detector. The current construction treats variables uniformly, with cross-variable information entering through these pooled responses.

Second, this multivariate evidence benefits detection. We progressively reduced the proportion of variables participating in aggregation.

Variables retained, nominal proportionTSB-AD-M VUS-PR (%)
100%50.28
75%49.18
50%47.65
25%45.09

Detection improves as more variables contribute evidence. Reducing coverage from 100% to 25% lowers VUS-PR by 5.19 percentage points, demonstrating the value of predictive responses from additional variables.

However, adding pooled evidence also causes substantial regressions on individual sequences. From the pooling ablations across all 180 TSB-AD-M Eval sequences, we selected two cases comparing variable-maximum-only detection with complete HieraNTP. Both configurations retrain the detector under the same fixed recipe.

Eval sequenceVariable maximum onlyHieraNTPVUS-PR decrease (pp)
SMAP 15368.913.0265.89
SMD 06834.6013.7720.83

VUS-PR is reported in percent. Figure 13 in Appendix F.6 compares evidence channels, split gain, final scores, and error locations across configurations. In SMD 068, adding variable-mean evidence increases false positives from 125 to 379 in the displayed window, introducing 254 additional false alarms, while both configurations detect all 101 anomalous points in that window. In SMAP 153, normal-region scores become more prominent relative to the anomaly, substantially degrading anomaly ranking. Both configurations miss the same 41 anomalous points at threshold 0.5.

These errors reflect the information compression introduced by fixed pooling. Means can dilute anomalous responses confined to a few variables, while maxima can emphasize irrelevant fluctuations in an individual variable. Aggregation also compresses information about which variables produced these responses. Adding pooled evidence changes the combinations learned by the detector and can consequently affect false alarms and anomaly ranking.

We also conducted the controlled relational-anomaly experiment suggested by the reviewer. Using real normal TSB-AD-M segments, joint and variable-specific circular shifts construct dependency-preserving and dependency-altered conditions. Each variable’s empirical marginal distribution remains unchanged, and four counterbalanced repeats ensure that both conditions use the same collection of complete univariate trajectories. Relation-task labels from 12 Tuning segments train a 32-tree extension of the shared base detector. Evaluation uses 12 different Eval sequences and the existing 4992-dimensional representation.

DetectorVUS-PR (%)VUS-ROC (%)
Shared base detector68.0358.47
HieraNTP, original anomaly labels67.8858.30
HieraNTP, relation-task adaptation80.2070.59

These metrics evaluate the constructed relational-anomaly intervals. Relation-task adaptation improves VUS-PR by 12.32 percentage points, showing that the representation retains cross-variable cues that supervision can exploit. Together with the failure cases, these findings motivate more explicit treatment of variable dependencies through relation learning and adaptive interactions before aggregation. We have incorporated the pooling failure analysis into Section 4.3, with variable-coverage and error results in Appendix E.4 (Tables 28–29), the controlled relation-task experiment in Appendix E.5 (Tables 30–32), and additional visualizations in Appendix F.6 (Figure 13).

Reviewer_1KMiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper identifies the predictive risk–decision gap in time-series anomaly detection (TSAD): deviations from learned temporal regularities do not necessarily correspond to task-defined anomalies, while prematurely aggregating heterogeneous predictive discrepancies into a scalar score may discard task-relevant information.

To address this gap, the authors argue for separating self-supervised temporal modeling from supervised anomaly decision-making, while preserving heterogeneous predictive evidence until task-specific interpretation. They propose HieraNTP, which pretrains a hierarchical tokenizer and next-token predictor to capture temporal regularities at multiple scales. The resulting predictive discrepancies are organized into a structured risk representation across token, local-statistic, and observation spaces. Finally, an XGBoost detector is trained with anomaly labels to learn task-specific decision rules.

Experiments on TSB-AD and TAB report substantial improvements over existing baselines. Controlled ablation studies further investigate the contributions of evidence preservation, temporal structure, and task-specific adaptation.

Critical strengths

1. Clear and well-motivated problem formulation. The predictive risk–decision gap provides a clear and meaningful motivation for the proposed framework. Figure 1 intuitively illustrates the problem and helps establish a coherent connection between the motivation and the overall design.

2. Strong empirical results and informative ablation studies. HieraNTP achieves substantial improvements on both TSB-AD and TAB. More importantly, controlled ablations under matched supervised detector training demonstrate the benefits of preserving heterogeneous predictive evidence over simple scalar aggregation, providing empirical support for the proposed information-preservation strategy.

Critical weaknesses

1. Insufficient isolation of the proposed methodological contributions. While the ablation studies demonstrate the benefits of preserving heterogeneous predictive evidence over simple scalar aggregation, they do not sufficiently isolate the contributions of the proposed hierarchical predictive backbone and structured risk representation. In particular, existing TSFM baselines are evaluated using their original detection protocols rather than comparable supervised detectors. Consequently, it remains unclear whether HieraNTP's improvements arise from its specific representation design or more generally from combining rich pretrained features with supervised anomaly detection.

2. Limited direct empirical support for the predictive risk–decision gap. Although the predictive risk–decision gap provides a compelling motivation, the experiments do not directly demonstrate how HieraNTP addresses the distinction between temporal regularity violations and task-defined anomalies. The ablations and feature-usage analyses support the usefulness of complementary evidence, but do not clearly establish whether the improvements arise from resolving this gap rather than from richer feature representations. Figure 1 illustrates the intended distinction, but remains schematic.

Questions for authors

1. Incremental contribution of HieraNTP's representation and predictive backbone Could the authors provide further evidence that the proposed hierarchical predictive backbone and structured risk representation offer benefits beyond existing pretrained time-series representations under matched downstream supervision? For example, comparisons using multiview evidence from TSPulse or another suitable TSFM, or simpler multiscale statistical features, followed by the same XGBoost detector and anomaly-label supervision, would help clarify the incremental contribution.

2. Direct validation of the predictive risk–decision gap. Could the authors provide more direct evidence that HieraNTP distinguishes temporal irregularities from task-defined anomalies? For instance, the Yahoo example in Figure 1 could be supported with actual predictive evidence and anomaly scores, showing how the structured representation and supervised detector distinguish regions A and B where scalar discrepancy scoring fails. Alternatively, a controlled analysis involving different anomaly criteria could help substantiate this claim. Convincing evidence addressing these concerns, particularly the incremental contribution of the proposed representation under matched supervision, could lead me to reconsider my recommendation.

2 comments

To W1Q1: Incremental Contribution of the Hierarchical Predictive Backbone and Structured Risk Representation under Matched Downstream Supervision

Thank you for suggesting this controlled comparison. We evaluated TSPulse, Chronos-Bolt, MOMENT, and a multiscale statistical baseline based on patch means and variances on TSB-AD. Following HieraNTP, we retain complementary prediction and reconstruction discrepancies and encoder representations, as available. Each method uses the same anomaly labels and XGBoost training pipeline, including a shared base detector and dataset-specific extensions.

Table R1 compares zero-shot detection, fixed evidence averaging, and task-adaptive detection on TSB-AD-U and TSB-AD-M. All supervised detectors, including HieraNTP, are retrained with seed 42. Zero-shot TSPulse and MOMENT results follow the TSB-AD leaderboard. Chronos-Bolt uses Bolt-base with a context length of 2048 and one-step mean-squared prediction error.

Table R1. VUS-PR (%) on TSB-AD-U / TSB-AD-M.

MethodZero-shotFixed evidence averagingTask-adaptive detection
TSPulse48 / 3632.98 / 30.4557.13 / 48.85
MOMENT38 / —28.44 / 31.9252.03 / 49.23
Chronos-Bolt29.24 / —33.01 / 29.6957.33 / 49.04
Multiscale statistics—37.83 / 37.3745.51 / 40.53
HieraNTP—43.65 / 35.2765.69 / 50.31

MOMENT uses eight reconstruction-error channels for fixed evidence averaging and 56 channels, including projected encoder representations, for task-adaptive detection.

Fixed evidence averaging underperforms the original zero-shot scoring for TSPulse and MOMENT on U. Averaging may dilute informative responses amid redundant or competing signals. All three TSFM alternatives fall below the statistical baseline under fixed evidence averaging. Evidence processing and task-adaptive detection improve VUS-PR by 73.2–82.9% on U and 54.3–65.2% on M over fixed evidence averaging, exceeding the statistical baseline. These results support the importance of the predictive risk–decision gap. Pretrained models provide informative responses to temporal irregularities, while the detector learns from anomaly labels how they correspond to the target anomaly criterion.

HieraNTP exceeds the strongest supervised TSFM alternatives by 8.36 percentage points on U and 1.08 on M. Its hierarchical predictive evidence and structured risk representation retain an advantage under comparable downstream supervision.

We have included this comparison in Section 4.3 (Table 4), with evidence construction, feature dimensions, averaging rules, and complete matched-supervision results documented in Appendix E.1 (Tables 23–24), and detector configurations in Appendix C.2.

To W2Q2: Direct Empirical Validation of the Gap Between Predictive Risk and Task-Defined Anomaly Decisions

Thank you for suggesting direct validation of the predictive risk–decision gap. Figure 5 in the revised manuscript presents predictive risk evidence, detector usage, and task-adaptive anomaly scores for the Yahoo example in Figure 1. The original annotations label region A, where local peaks disappear, as normal and region B, where the cyclic pattern is disrupted, as anomalous.

HieraNTP retains evidence responding to different violations of temporal regularity. A1–A3, drawn from local-statistic, token-space, and observation-space discrepancies, respond strongly to A. For example, A1 achieves a window VUS-PR of 0.999 under a post-hoc diagnostic criterion that treats A as positive, but only 0.300 under the original Yahoo labels. Meanwhile, B1–B3 capture patch-variance discrepancies at patch length 8 and respond more strongly to B. These observations show how a substantial predictive discrepancy can identify a temporal irregularity that the target task regards as normal.

The trained task-adaptive detector uses these responses selectively. The Yahoo-specific extensions allocate no splits to A1–A3 and make substantial use of B1–B3, with B1 accounting for 46.5% of the extensions’ total split gain. B1’s mean local contribution to the anomaly logit is −0.1695 in A and +0.0578 in B, lowering the score in the normal region and increasing it in the anomalous region. The complete detector achieves a window VUS-PR of 0.892 under the original Yahoo labels. Fixed evidence averaging applies a uniform aggregation rule, while task-adaptive detection learns evidence combinations from training labels. This analysis uses the original trained detector and Yahoo annotations.

Anomaly semantics depend on the application context. Increased server utilization may reflect ordinary workload, whereas an unexpected rise in process temperature may require intervention. Similarly, a flat zero response may indicate an offline device, while a brief zero traffic count may correspond to a red light. This dependence on task semantics distinguishes anomaly detection from temporal prediction and motivates HieraNTP to retain multiview predictive risk evidence and learn its selection and combination from target-task anomaly labels.

We have included this empirical counterpart to Figure 1 in Section 4.5 (Figure 5) and documented the channel definitions, retrospective channel selection, local contributions, and evaluation settings in Appendix F.3 (Table 35).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.