acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Encoder Learning in Time-Series Anomaly Detection

Abstract

Time-series anomaly detection (TSAD) commonly learns representations from historical observations to distinguish normal from anomalous behavior. However, in the absence of anomaly supervision, learning alone faces a fundamental limitation: its objective does not directly specify which aspects of the data are actually relevant for anomaly detection. Historical observations, meanwhile, provide a concrete basis for evaluating new observations. This raises a fundamental question: if historical observations can serve directly as empirical references, is encoder learning really necessary for strong anomaly detection? As a direct test of this question, we introduce Fixed Anomaly Representation with Observed References (FARO), a deliberately simple architecture with no encoder training. FARO fixes a randomly initialized encoder, stores historical features in channel-specific reference banks, and scores new observations by their deviation from nearest historical references. The resulting channel-wise deviations support both anomaly detection and localization. Despite its simple design, FARO achieves state-of-the-art anomaly detection performance on TSB-AD-U/M and competitive channel localization on SMD, HAI, and SWaT. More importantly, controlled experiments show that, with historical references available, encoder training provides no consistent additional gain and can even degrade detection. This pattern also holds across multiple existing TSAD backbones. Together, these findings challenge the default reliance on encoder learning in TSAD and highlight empirical references as a strong alternative for exploiting historical observations. The source code is available at AnonymizedURL.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NnAGOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether target-specific encoder learning is necessary when historical observations are available as empirical references. FARO uses a fixed, randomly initialized time–frequency encoder, stores channel-specific features of historical observations, and scores each query based on its distance to nearby references. The same channel-level deviations are used for anomaly detection and localization. Experiments show that FARO is competitive and that the evaluated encoder-training procedures do not consistently improve mean performance.

Critical strengths
  1. Practical performance: Table 1 and Appendix C report strong detection performance across both benchmark tracks using ten independently initialized FARO encoders. The comparisons include learned, pretrained, and random-feature detectors rather than only a weak baseline.
  2. Interesting findings on the role of the encoder: The finding that a fixed, randomly initialized encoder can perform competitively is interesting because it challenges the common expectation that encoder learning is necessary for anomaly detection.
Critical weaknesses
  1. Dependence on historical data: TSFMs can be useful when only limited target-specific training data are available, and few-shot learning can sometimes provide satisfactory results. Since FARO directly relies on historical data for its reference set, its performance may deteriorate when these data are limited.
  2. Generality of encoder comparisons: The TSFM comparisons in Table 1 evaluate complete detectors rather than pretrained representations under the same reference-scoring procedure. The referee would like to see controlled comparisons across representative encoder architectures, including randomly initialized, frozen pretrained, and fine-tuned variants.
Questions for authors
  1. How does FARO perform when only a limited amount of historical data is available? Can the authors evaluate progressively shorter training prefixes, construct references using only each prefix, and keep the test segment fixed? Comparing FARO with pretrained TSFMs under the same target-history budgets would clarify how much historical data FARO needs to remain competitive.
  2. Can the authors extend the encoder comparisons to representative architectures, including TSFMs? For each suitable backbone, please compare randomly initialized and frozen, pretrained and frozen, and pretrained and fine-tuned variants, using the same feature readout, reference data, and scoring procedure.
4 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank the reviewer for their thoughtful feedback. To address the two questions raised, we conducted additional experiments and present the results below.

Q1. Limited Historical Data and Comparison with Pretrained TSFMs

“How does FARO perform when only a limited amount of historical data is available? ... Comparing FARO with pretrained TSFMs under the same target-history budgets would clarify how much historical data FARO needs to remain competitive.”

FARO outperforms the compared fine-tuned TSFM detectors even with limited historical data, while benefiting from additional history.

We evaluated FARO and four pretrained TSFMs (TSPulse, MOMENT, TimesFM, and Chronos) with 5%, 10%, 25%, 50%, 75%, and 100% of the training history. At each budget, FARO constructs its reference bank using only the available training prefix, and each TSFM is fine-tuned using that same prefix. The test segment is fixed across all methods and budgets.

Since repeatedly fine-tuning four TSFMs across six history budgets incurs substantial computational cost, we used a fixed, performance-independent stratified subset of 50 series (32 U, 18 M), covering 17 U and 14 M dataset families. Selection accounted for sequence length and channel count, and every method was evaluated on the same series.

Table 1(a). Limited-history detection on TSB-AD-U (VUS-PR ↑)

Model5%10%25%50%75%100%
FARO0.44690.46590.54170.52840.53340.5420
TSPulse-FT0.34080.35570.39950.45310.46330.4530
MOMENT-FT0.23850.23720.23490.23630.23940.2429
TimesFM-FT0.20310.20890.20360.20460.21010.2096
Chronos-FT0.17370.17550.17550.17530.17480.1753

Table 1(b). Limited-history detection on TSB-AD-M (VUS-PR ↑)

Model5%10%25%50%75%100%
FARO0.32580.32800.36160.39010.40260.4101
TSPulse-FT0.31370.32230.32610.33260.33030.3317
MOMENT-FT0.24320.24850.24830.24890.24950.2496
TimesFM-FT0.20450.20130.20190.20240.20000.2038
Chronos-FT0.18170.18350.18320.18280.18300.1826

FARO achieves the highest mean VUS-PR at every history budget on both tracks. With only 5% history, FARO reaches 0.4469 (U) and 0.3258 (M). On U, 25% history already achieves nearly the full-history performance (0.5417 vs. 0.5420). On M, performance rises from 0.3258 to 0.4101 as history increases from 5% to 100%.

These results show that FARO does not rely on extensive historical data to achieve strong anomaly detection, remaining effective even with limited history. Moreover, retaining more historical observations as references can further improve detection performance.

AuthorsOpenReview ↗

Official Comment by Authors

Q2. Generality of Encoder Comparisons

“For each suitable backbone, please compare randomly initialized and frozen, pretrained and frozen, and pretrained and fine-tuned variants, using the same feature readout, reference data, and scoring procedure.”

Pretraining improves detection across all evaluated backbones, whereas additional fine-tuning provides no consistent gain once historical references are available.

We evaluated TSPulse, MOMENT, TimesFM, and Chronos under three conditions: Random Frozen, Pretrained Frozen, and Fine-tuned. Within each backbone, we kept the feature readout, reference/query windows, normalization, reference bank, and cosine 3-NN scoring identical across conditions. This directly separates the contribution of pretraining from that of subsequent fine-tuning. The experiment covers the complete TSB-AD-U/M benchmarks (350 U and 180 M series).

Table 2(a). Controlled encoder comparison on TSB-AD-U (VUS-PR ↑)

BackboneRandom FrozenPretrained FrozenFine-tuned
TSPulse0.18380.31500.3144
MOMENT0.26470.38940.3728
TimesFM0.20740.25660.2552
Chronos0.24920.36980.3675

Table 2(b). Controlled encoder comparison on TSB-AD-M (VUS-PR ↑)

BackboneRandom FrozenPretrained FrozenFine-tuned
TSPulse0.19790.34720.3419
MOMENT0.28790.33390.3328
TimesFM0.20040.24940.2493
Chronos0.24270.32610.3276

First, Pretrained Frozen outperforms Random Frozen in all eight backbone–track comparisons, demonstrating the value of pretrained representations for reference-based anomaly detection.

Second, fine-tuning reduces mean VUS-PR in seven of eight comparisons; the only improvement is +0.0015 for Chronos on M.

These results distinguish the benefit of pretraining from the additional benefit of fine-tuning. Even when pretrained representations improve detection, further fine-tuning does not consistently help when historical observations are retained as references. This strengthens our central argument that the incremental value of encoder learning should be examined relative to the evidence already available in historical references.

 

We thank Reviewer NnAG for suggesting these valuable experiments. We will incorporate the main results into the revised manuscript and provide detailed experimental settings and additional analyses in the Appendix.

Thank you for the additional experiments, which substantially address my concerns about limited historical data and encoder generality. The results support the value of pretraining while showing no consistent additional benefit from target-specific fine-tuning under the tested settings. Please make this distinction explicit in the manuscript’s main claims and include the absolute history lengths and key experimental details in the revision. I maintain my weak acceptance recommendation.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you very much for your thoughtful feedback and encouraging assessment.

In response to your suggestion, we have added the absolute history lengths and clarified the history-budget definition in Appendix H.2.

We sincerely appreciate your careful evaluation and constructive suggestions.

Reviewer_ferWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether encoder training improves time-series anomaly detection when historical observations are retained as references. It proposes FARO, which combines a fixed random time–frequency encoder with channel-specific reference banks and nearest-neighbor scoring for detection and channel localization. Experiments on TSB-AD-U/M and three localization datasets report competitive performance, while controlled comparisons across multiple backbones and training objectives find no consistent additional benefit from encoder training under the evaluated reference-based settings.

Critical strengths
  1. The controlled comparisons across multiple backbones and training objectives are a substantive strength. By matching historical references and scoring procedures between trained and untrained encoders, the paper provides useful evidence about the additional value of encoder learning. FARO’s strong results on both complete TSB-AD tracks also establish a practical baseline without encoder training.
Critical weaknesses
  1. Limited conceptual advance beyond existing reference-based detection. FARO combines fixed random features with historical nearest-neighbor scoring, while the paper already identifies related random-feature and historical-reference methods such as RandomAD and PaAno. Its main contribution therefore rests on the insight gained from the training comparisons. However, the paper primarily establishes an empirical observation without sufficiently explaining why training offers limited additional benefit or distinguishing the effects of representation design, historical coverage, and scoring strategy. This limits the depth of the claimed contribution of “rethinking encoder learning.”
  2. Insufficient characterization of when encoder learning is beneficial. The conclusions rely largely on aggregate benchmark performance, which can conceal differences across anomaly types and data regimes. Training does improve some evaluated settings, including multivariate CATCH and PaAno, but these cases are not examined in depth. Moreover, the FARO-specific training controls mainly use 50–100 updates, with longer trajectories evaluated on a limited subset. The evidence supports the usefulness of fixed encoders under the tested configurations, but does not adequately establish the conditions governing the value of training. Without this characterization, the broader methodological implications remain uncertain.
Questions for authors
  1. Clarification of contribution: Beyond demonstrating competitive performance without encoder training, what specific new insight does the paper establish relative to existing random-feature and historical-reference detectors? Which experiments distinguish the contribution of encoder learning from that of the representation design and reference-based scoring?

  2. Clarification and supporting evidence: Under what conditions does encoder training help or hurt? Can the authors explain the gains observed for multivariate CATCH and PaAno and provide a breakdown by anomaly type or data regime? How sensitive are the FARO training comparisons to longer optimization budgets beyond the selected 50–100 updates?

5 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank the reviewer for the constructive feedback. We conducted additional experiments to clarify our contribution and examine the role of historical references in encoder learning.

Q1. Novelty and the Contribution of Encoder Learning

“Beyond demonstrating competitive performance without encoder training, what specific new insight does the paper establish ...? Which experiments distinguish the contribution of encoder learning from that of the representation design and reference-based scoring?”

Our contribution is to isolate and characterize the incremental value of encoder learning when historical observations are directly used as empirical references. Our experiments reveal a key insight: this additional benefit generally diminishes as the reference budget increases.

RandomAD uses random kernels with data-driven feature/scale selection; PaAno combines learned patch representations with historical-reference scoring. In contrast, FARO is deliberately training-free and tests what encoder learning adds when historical observations are already available as empirical references.

1. Disentangling Representation, References, Scoring, and Encoder Learning

Table 6 and Appendix E separately examine representation design, reference sources, and scoring. On TSB-AD-M, time-frequency features outperform temporal-only features (0.485 vs. 0.409), own-channel references outperform pooled references (0.485 vs. 0.464), and k-NN outperforms centroid scoring (0.485 vs. 0.411).

We also compare native trained detectors with random and trained reference-based variants across existing backbones, distinguishing the gains from direct historical comparison from those of encoder training under the same reference setting.

Table 1(a). Existing TSAD backbones on TSB-AD-M (VUS-PR ↑)

BackboneTrained NativeRandom + Ref.Trained + Ref.
KAN-AD0.3340.4300.431
ModernTCN0.3460.4350.433
ScatterAD0.2400.3310.329
CATCH0.3300.4050.412

Table 1(b). Existing TSAD backbones on TSB-AD-U (VUS-PR ↑)

BackboneTrained NativeRandom + Ref.Trained + Ref.
KAN-AD0.3190.3980.393
ModernTCN0.3200.4760.476
ScatterAD0.2590.3790.384

Reference-based variants use matched references and scoring (original Table 3).

In all seven backbone–track comparisons, Random + Reference outperformed the original trained detector. Under matched reference-based detection, encoder learning provided no consistent additional gain (Table 3), consistent with FARO's five-objective controls (Table 5). Thus, directly retaining and comparing historical observations can matter more than learning encoder weights.

2. New Finding: Reference Availability and the Value of Encoder Learning

To directly address the reviewer's question about disentangling encoder learning from representation design and reference-based scoring, we conducted new controlled experiments across five learning objectives and five reference budgets (32, 128, 512, 1,024, and 3,000) on the full TSB-AD-U/M benchmarks. Fixed and trained encoders use identical architectures, reference windows, and scoring procedures, isolating the incremental contribution of encoder learning.

Table 2. Reference budget and incremental encoder-learning gain (VUS-PR ↑)

TrackCondition32 refs3,000 refs
UFixed encoder0.48880.5705
ULearning gain — One-class+0.0328−0.0208
MFixed encoder0.42990.4817
MLearning gain — InfoNCE+0.0398+0.0036
MLearning gain — Forecasting+0.0327+0.0201
MLearning gain — One-class−0.0332−0.0260

Learning gain = Trained − Fixed. Representative endpoints; full results in the Appendix.

The incremental learning gain decreased in 9 of 10 objective–track comparisons as the reference budget grew. For U One-class, the gain reversed from +0.0328 to −0.0208; for M InfoNCE, it dropped from +0.0398 to +0.0036. Three comparisons also showed negative target- and family-level 95% bootstrap confidence intervals.

Meanwhile, the fixed encoder improved from 0.4888 to 0.5705 (U) and from 0.4299 to 0.4817 (M). At 3,000 references, fixed encoders outperformed trained encoders in 8 of 10 conditions.

These results identify reference availability as a key condition governing the marginal value of encoder learning. When references are scarce, learned features may help capture similarities among observed patterns. As the bank grows, queries can be compared with a richer set of actual historical behaviors, leaving less additional benefit from training in most evaluated settings.

 

Our study goes beyond a strong fixed-random baseline by showing how the incremental value of encoder learning changes with reference availability. TSAD research should evaluate what learning adds after historical observations are retained and directly exploited, rather than assume encoder training is necessary.

AuthorsOpenReview ↗

Official Comment by Authors

Q2. Conditions Under Which Encoder Learning Helps or Hurts

“Under what conditions does encoder training help or hurt? Can the authors explain the gains observed for multivariate CATCH and PaAno and provide a breakdown by anomaly type or data regime?”

We analyzed CATCH and PaAno through dataset-level comparisons, BatchNorm controls, and anomaly-type analysis, finding that their training gains depend on data regime and normalization.

1. CATCH: Training Gains Concentrated in ECG Series

CATCH's +0.0072 mean gain on TSB-AD-M is concentrated in 39 MITDB/SVDB ECG series.

Table 3. Dataset-wise encoder-learning effect in CATCH (VUS-PR ↑)

Dataset groupRandom + Ref.Trained + Ref.Gain
All M1800.4050.412+0.007
MITDB/SVDB (39 ECG series)0.1550.226+0.071
Remaining 141 series0.4740.464−0.010

All 39 are two-channel ECG series with long histories, recurring normal waveforms, and concentrated spectral energy. Before training, CATCH assigned high scores to recurring, high-amplitude normal ECG patterns. Training reduced these scores, while anomalous-timestamp score percentiles rose from 67.76% to 75.57%.

This suggests that CATCH's time-frequency reconstruction objective helps distinguish recurring normal ECG patterns from actual anomalies. Excluding the ECG series reverses the mean training gain to −0.0105, showing that the overall gain is ECG-driven.

2. PaAno: BatchNorm Effects and Benefits on Local Patterns

PaAno's native training procedure yields a +0.0197 VUS-PR gain on TSB-AD-M. To isolate weight learning, we compared initial and trained encoders with matched references and common BatchNorm recalibration.

Table 4. Incremental encoder-learning effect in PaAno (M180, VUS-PR ↑)

ConditionInitial (BN-only)TrainedGain
Native procedure0.40740.4270+0.0197
Common BN recalibration0.40760.4086+0.0011

With common BN recalibration, the incremental gain from encoder-weight training falls to +0.0011. This indicates sensitivity to the interaction between learned representations and BatchNorm statistics.

Training still benefits certain datasets. On Exathlon, performance improves on 22 of 25 series, with a mean VUS-PR gain of +0.0205 even after BN recalibration. Exathlon contains recurring Spark streaming workloads with local deviations caused by resource contention and process failures. PaAno's patch-level representation learning may help distinguish these deviations from recurring normal workload patterns.

Taken together, CATCH's gains are concentrated where recurring ECG waveforms align with its time-frequency reconstruction objective, whereas PaAno's average gain is normalization-sensitive but persists on Exathlon. These results explain why encoder learning can help in specific data regimes without providing consistent gains when historical references are available.

3. Encoder-Learning Effects by Anomaly Type

We analyzed 12,196 anomaly intervals in TSB-AD-M, using anomaly-type definitions informed by Figure 3 and prior work [1, 2]. Multiple anomaly-type labels per series were allowed.

Table 5. Encoder-learning gain by anomaly type (ΔVUS-PR = Trained − Initial)

Anomaly typeSeriesIntervalsCATCHPaAno
Global Point10810,490+0.0166+0.0009
Context17114+0.0184−0.0211
Shapelet2031+0.0131−0.0091
Collective59201+0.0070−0.0132
Trend38106+0.0066−0.0211
Mixed/Seasonal57164−0.0061+0.0002
Uncertain681,090+0.0253−0.0151
Overall M18018012,196+0.0072+0.0011

CATCH gains +0.0166 on series containing global-point anomalies, but this changes to −0.0100 after excluding ECG series. This indicates that the apparent anomaly-type gain is more closely associated with ECG-specific characteristics.

PaAno shows a −0.0211 performance change on both Context and Trend series. For Trend, performance declines on 28 of 38 series, accompanied by reduced coverage of labeled anomaly intervals. This suggests that learning local patch patterns may be less effective for context- or trend-dependent anomalies.

The anomaly-type breakdown reveals no category with consistent improvements across both models.

Implications

The positive average gains for CATCH and PaAno reflect specific data characteristics and normalization effects. When historical observations are directly available as references, additional encoder training does not guarantee better anomaly detection, reinforcing the central finding of our study.

References

[1] Lai et al. (2021). Revisiting Time Series Outlier Detection: Definitions and Benchmarks. NeurIPS Datasets and Benchmarks.

[2] Blázquez-García et al. (2021). A Review on Outlier/Anomaly Detection in Time Series Data. ACM Computing Surveys.

AuthorsOpenReview ↗

Official Comment by Authors

Q2 (Continued). Longer Optimization and Feature Geometry

“How sensitive are the FARO training comparisons to longer optimization budgets beyond the selected 50–100 updates?”

To examine whether the limited gains in our training comparisons were due to insufficient optimization, we extended training to 1,000 updates across all 530 series in TSB-AD-U/M, covering five learning objectives.

1. Longer Optimization on the Full Benchmark

We evaluated Reconstruction, InfoNCE, Triplet, Forecasting, and One-class objectives at 100, 400, 600, and 1,000 updates. Fixed and trained encoders used matched initializations, identical historical references, query windows, and anomaly scoring procedures.

Table 6. Extended encoder training on TSB-AD-U/M (ΔVUS-PR = Trained − Fixed)

TrackObjective1004006001,000
UReconstruction+0.0160+0.0032−0.0167−0.0269
UInfoNCE−0.0413−0.1293−0.1262−0.1365
UTriplet−0.0229−0.0522−0.0538−0.0600
UForecasting+0.0175−0.0031−0.0309−0.0630
UOne-class−0.0417−0.0198−0.0195−0.0207
MReconstruction+0.0146+0.0113+0.0059−0.0066
MInfoNCE−0.0038−0.0070+0.0006+0.0036
MTriplet−0.0178+0.0025−0.0027−0.0022
MForecasting+0.0153+0.0042+0.0166+0.0201
MOne-class−0.0576−0.0270−0.0251−0.0260

At 1,000 updates, all five objectives on U underperformed the matched fixed encoder. On M, three of the five objectives also failed to provide additional gains.

In particular, U Reconstruction and Forecasting initially showed gains of +0.0160 and +0.0175 at 100 updates, respectively, but these gains declined and eventually became negative with further optimization.

In contrast, M Forecasting maintained a positive gain throughout training, reaching +0.0201 at 1,000 updates. However, this was an isolated result among the evaluated objective–track combinations, and longer optimization did not yield consistent improvements over the fixed encoder.

2. Feature-Geometry Analysis

To further understand the effects of extended training, we analyzed training loss, distances to historical references, and nearest-neighbor structures.

Table 7. Feature geometry and detection performance after 1,000 updates

ObjectiveLoss reductionNormal–anomaly distance gap ΔTop-3 neighbor overlapΔVUS-PR
U InfoNCE9.74%+0.03126.1%−0.1365
U Forecasting8.61%+0.01828.2%−0.0630
M Forecasting8.72%+0.003338.4%+0.0201

Distance gap: difference between median anomaly and normal distances to historical references. Neighbor overlap: fraction of shared top-3 nearest neighbors before and after training.

For U InfoNCE, training loss decreased by 9.74%, and the normal–anomaly distance gap increased by +0.0312. The Top-3 neighbor overlap was only 6.1%, indicating a substantial change in the nearest-neighbor structure.

Despite these changes, VUS-PR decreased by 0.1365. U Forecasting showed a similar pattern: reduced training loss and an increased distance gap, but lower detection performance.

In contrast, M Forecasting retained a higher neighbor overlap of 38.4% and achieved a VUS-PR gain of +0.0201.

These results show that encoder learning can substantially alter representations and their relationships with historical references without consistently improving detection. In particular, the U InfoNCE results demonstrate that substantial changes in the reference-neighborhood structure do not necessarily translate into better anomaly detection.

Implications

The extended experiments across the full benchmark show that the limited additional benefit of encoder learning cannot be attributed solely to a short optimization budget.

When historical observations are retained as empirical references, encoder learning can provide additional gains in specific settings, but even extended training does not consistently outperform the matched fixed encoder. This further supports our central finding that strong anomaly detection can be achieved by directly exploiting historical references without making encoder learning a prerequisite.

We will incorporate the additional experiments and key analyses into the revised manuscript, with detailed settings and full results provided in the Appendix.

Reviewer_8bHPOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studied on encoder learning in Time-Series Anomaly Detection (TSAD). The authors introduce FARO (Fixed Anomaly Representation with Observed References), a simple architecture that uses a randomly initialized, fixed time-frequency encoder. It stores historical features in channel-specific reference banks and scores new observations based on their distance to the historical references. Experiments on TSB-AD-U/M and localization evaluation demonstrate that FARO achieves good performance,.

Critical strengths
  1. The paper is clear and easy to follow.
  2. time-frequency encoder follows existing modeling and is reasonable.
Critical weaknesses
  1. Limited Methodological Novelty and nothing new from this paper: While the empirical experiments that "encoder learning is unnecessary for search based Time-Series Anomaly Detection" is conducted, the proposed FARO method itself is essentially a nearest-neighbor search in a random-mapping feature space (the same to RandomAD, memory-based methods or KNN). The paper is an empirical study that existing works have already studied, rather than a novel contribution expected at ICLR.
  2. Vulnerability to Concept Drift: The core assumption of FARO is that historical observations remain representative of "normal" behavior. The method lacks adaptive reference management. In real-world, non-stationary environments where normal operating regimes evolve (concept drift), a static reference bank will inevitably trigger massive false alarms. The failure to address this fundamental limitation of memory-based methods severely restricts the practical utility of the proposed framework.
  3. encoder learning is still necessary for reconstruction and forecasting based Time-Series Anomaly Detection, and this paper still fails on some datasets.
Questions for authors

See Weaknesses

9 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank the reviewer for the constructive comments and valuable suggestions. We address each point below and will revise the manuscript accordingly.

Q1. Novelty and Empirical Insights Beyond Existing TSAD Methods

Our contribution is to isolate and characterize the incremental value of encoder learning when historical observations are directly used as empirical references. Our experiments reveal a key insight: the additional benefit of encoder learning generally diminishes as the reference budget increases. This addresses our research question in the Introduction:

“If historical observations can be directly referenced and compared, is encoder learning really necessary for strong anomaly detection?”

Much of TSAD research seeks better representations of historical data. We instead ask what encoder learning adds when that same history is directly retained as empirical references. RandomAD uses random kernels and data-driven feature/scale selection; PaAno combines learned patch representations with reference scoring. We go beyond constructing a detector by disentangling the roles of historical references and encoder learning.

We thus present two complementary lines of evidence: a new controlled reference-budget experiment (Finding 1) and the cross-backbone comparisons reported in our original manuscript (Finding 2).

Finding 1. The incremental benefit of encoder learning decreases as the reference budget grows in most evaluated settings.

We conducted controlled experiments with reference budgets of 32 → 128 → 512 → 1,024 → 3,000. Fixed and trained encoders use the same architecture, reference windows, queries, and anomaly scoring; only encoder training differs.

Table A. Encoder-learning gain across reference budgets (ΔVUS-PR = Trained − Fixed)

TrackObjective32 refs3,000 refs
UOne-class+0.0328−0.0208
MInfoNCE+0.0398+0.0036
MForecasting+0.0327+0.0201
MOne-class−0.0332−0.0260

Results for all five budgets and 10 objective–track comparisons will be included in the Appendix.

The incremental training gain decreased in 9/10 objective–track comparisons as the reference budget grew from 32 to 3,000. For U/One-class, the gain reversed from +0.0328 to −0.0208; for M/InfoNCE, it shrank from +0.0398 to +0.0036. For U/One-class, U/InfoNCE, and M/InfoNCE, both target-level and family-level 95% confidence intervals for this change were negative. At 3,000 references, fixed encoders outperformed trained encoders in 8/10 comparisons. Meanwhile, increasing references alone improved fixed-encoder VUS-PR from 0.4888 to 0.5705 (U) and 0.4299 to 0.4817 (M).

The key insight is not merely that more references improve detection, but that the incremental benefit of encoder learning tends to diminish as reference availability increases.

Finding 2. Across existing TSAD backbones, historical references substantially improve detection, while encoder learning provides no consistent additional gain.

Original Table 3 compares each backbone's trained native detector with random and trained variants using historical references.

Table B. Native versus reference-based detection (VUS-PR ↑; original Table 3)

BackboneTrained NativeRandom + Ref.Trained + Ref.Learning Gain with Ref.
KAN-AD (M)0.3340.4300.431+0.001
ModernTCN (M)0.3460.4350.433−0.002
ScatterAD (M)0.2400.3310.329−0.002
CATCH (M)0.3300.4050.412+0.007
KAN-AD (U)0.3190.3980.393−0.005
ModernTCN (U)0.3200.4760.4760.000
ScatterAD (U)0.2590.3790.384+0.005

Learning Gain with Ref. = Trained + Ref. − Random + Ref.

First, Random + Reference outperformed each backbone's original trained detector in all seven comparisons. Direct historical comparison can therefore outperform native detection even with random encoder weights. Second, once historical observations are directly exploited as references, encoder learning provides no consistent additional gain (ΔVUS-PR: −0.005 to +0.007). FARO’s five-objective controls (Table 5) and PaAno’s training comparison (Table 4) further support this pattern.

These findings go beyond demonstrating a strong random-feature nearest-neighbor detector. We isolate the contribution of encoder learning, show how it changes with the reference budget, and establish across existing backbones that leveraging historical references can matter more than training the encoder.

Our broader implication is that TSAD research should evaluate not only how to learn from history, but also how to retain and directly exploit it—and what encoder learning adds once that evidence is available. FARO provides a strong training-free baseline for this question. We will incorporate the new findings into the main text and provide the full reference-budget results in the Appendix.

AuthorsOpenReview ↗

Official Comment by Authors

Q2. Vulnerability to Concept Drift and Adaptive Reference Management

We agree that evolving normal operating conditions can make static reference banks less representative, leading to increased false alarms. This limitation is already discussed in Section 5, where we identify adaptive reference management as an important research direction.

To directly address this concern, we conducted additional concept drift experiments and found that updating historical references can substantially mitigate drift-induced performance degradation without encoder retraining.

 

Controlled Concept Drift Experiment.

We constructed synthetic multivariate time series with controlled amplitude and inter-channel phase shifts, while preserving anomaly locations and relative anomaly magnitudes. We evaluated 10 time-series variants under no-drift and drift conditions, using three random initialization seeds for each trained model.

We also introduced a simple FIFO-based reference update to FARO, sequentially replacing the oldest references with features from recent observations while keeping the encoder and anomaly scoring unchanged.

Table C. Concept drift and adaptive reference management

MethodNo-drift VUS-PR ↑Drift VUS-PR ↑Drift FPR ↓
FARO0.86390.414619.03%
FARO (FIFO)0.83770.88762.03%

First, the results confirm the vulnerability of static reference banks. Under drift, FARO exhibits a substantial VUS-PR drop from 0.8639 to 0.4146, with an FPR of 19.03%.

Second, FIFO-based reference updates substantially restore detection performance. FARO (FIFO) achieves a drift VUS-PR of 0.8876, reduces FPR to 2.03%, and improves point recall from 78.39% to 100%. These improvements require neither encoder retraining nor changes to anomaly scoring.

Comparison with Existing TSAD Methods.

We further compared FARO with existing TSAD methods under the same drift conditions.

Table D. Comparison of TSAD methods under concept drift

MethodVUS-PR (No Drift → Drift)Drift FPR ↓Drift Recall ↑
FARO (FIFO)0.8377 → 0.88762.03%100%
FARO0.8639 → 0.414619.03%78.39%
PaAno (FIFO)0.5577 → 0.20606.97%71.04%
PaAno0.5420 → 0.1584100%100%
KAN-AD0.7141 → 0.072613.87%33.63%
CATCH0.0297 → 0.02700.99%0.35%

Results across 10 synthetic time-series variants. Learning-based methods were evaluated over three initialization seeds.

FARO (FIFO) achieves the highest drift VUS-PR among the compared methods. While FIFO also reduces false alarms for PaAno, its detection performance remains substantially below FARO. KAN-AD and CATCH likewise exhibit limited detection performance under the evaluated drift conditions.

 

Supplementary Real-World Evaluation.

To further validate reference adaptation on real-world data, we evaluated SMD and SWaT, two widely used TSAD benchmarks that have also been employed in studies addressing concept drift. We evaluated all 28 SMD machine series and SWaT, keeping the encoder fixed while retaining half of the initial reference bank and updating the remaining half via FIFO.

On SMD, mean VUS-PR improved from 0.3765 to 0.4215, while FPR decreased from 30.62% to 13.31%. On SWaT, VUS-PR improved from 0.3193 to 0.3759, while FPR decreased from 11.91% to 6.02%. Both datasets show improved detection performance and reduced false alarms through reference updates alone.

These results demonstrate that, even when static reference banks are vulnerable to concept drift, appropriately updating historical references can accommodate evolving normal behavior and mitigate this practical limitation.

Importantly, adapting to changing operating conditions does not necessarily require encoder retraining. Maintaining and updating empirical references can itself provide an effective mechanism for adapting anomaly detection to evolving normal behavior.

These results provide direct empirical support for adaptive reference management, as discussed in Section 5, and highlight the importance of maintaining representative historical references in TSAD.

We will incorporate the main results into the revised manuscript, with detailed experimental settings and additional results provided in the Appendix.

AuthorsOpenReview ↗

Official Comment by Authors

Q3. On the Necessity of Encoder Learning in TSAD

"encoder learning is still necessary for reconstruction and forecasting based Time-Series Anomaly Detection, and this paper still fails on some datasets."

Reconstruction and forecasting are well-established approaches in TSAD. However, the usefulness of learning reconstruction or forecasting objectives does not establish that encoder learning is necessary for strong anomaly detection.

Our research question is not whether encoder learning is universally unnecessary, but what additional benefit encoder learning provides when historical observations can be directly exploited as empirical references.

To investigate this, we conducted controlled comparisons across existing TSAD backbones (Table 3), five learning objectives applied to FARO (Table 5), and PaAno, which already employs historical-reference-based detection (Table 4). Across these experiments, encoder learning provides no consistent additional gain in anomaly detection when historical references are available.

Moreover, our evidence is not limited to a small selection of time series. FARO was evaluated on all 530 time series in TSB-AD-U/M, spanning diverse domains including industrial systems, IT infrastructure, and physiological signals. FARO achieved the highest performance among the compared methods across all six detection metrics on both benchmark tracks (Table 1). These results demonstrate that historical references can support strong anomaly detection without encoder learning across diverse real-world time series.

Our central message is not that encoder learning should be discarded. Rather, when historical observations can serve directly as anomaly evidence, the additional detection value of encoder learning should be evaluated rather than assumed.

Taken together, our results highlight that how historical observations are retained, compared, and updated is as important a research question in TSAD as how their representations are learned.

We will further clarify the scope and implications of our research question in the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Updated Manuscript: Summary of Revisions

We sincerely thank all reviewers for their constructive feedback. We have uploaded the revised manuscript, with revisions highlighted in blue. The major changes are summarized below.

Major Changes in the Revised Manuscript

In response to Reviewer 8bHP:

  • Novelty and contribution (Introduction, Section 4.4, Figure 4): Clarified our central contribution and added reference-budget experiments showing that the incremental benefit of encoder learning decreases in 9 of 10 comparisons as more historical references become available.
  • Concept drift (Section 5, Table 8; Appendix A.4): Added synthetic and real-world experiments on SMD and SWaT, demonstrating that FIFO reference updates can improve detection without encoder retraining.
  • Scope of our findings (Introduction, Section 5, Conclusion): Clarified that our findings concern the additional value of encoder learning when historical references are available.

In response to Reviewer ferW:

  • Controlled analyses (Sections 4.3–4.5, Figure 4): Strengthened comparisons of encoder learning, reference availability, feature representations, and scoring strategies across different architectures and learning objectives.
  • Training gains and exceptions (Appendix G.3, Table 25): Added dataset-specific and anomaly-type analyses of CATCH and PaAno, including the effects of BatchNorm recalibration.
  • Longer optimization (Section 4.4, Figure 5; Appendix D.2): Extended training to 1,000 updates across five objectives on TSB-AD, showing that lower training loss does not consistently improve detection.

In response to Reviewer NnAG:

  • Limited historical data (Section 4.3; Appendix H.2, Table 26): Added comparisons with four fine-tuned TSFM detectors across six history budgets (5–100%), showing FARO's strong performance even with limited history.
  • Controlled TSFM comparisons (Section 4.3, Table 5; Appendix H.1): Added random frozen, pretrained frozen, and fine-tuned encoder comparisons across four TSFMs on the complete TSB-AD-U/M benchmarks, distinguishing the benefits of pretraining from additional target-specific learning.

These revisions further support our central finding that historical observations can serve as strong empirical references for anomaly detection, while additional encoder learning provides no consistent performance gain when such references are available.

We appreciate the reviewers' feedback and hope the revised manuscript addresses their concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.