acceptodds
Under review as a conference paper at ICLR 2027

From Forecasts to Anomaly Scores: Reusing Frozen Time Series Foundation Models

Abstract

Frozen forecasting time series foundation models (TSFMs), when used for anomaly detection based on one-step forecast error (conventional scoring), have been reported to perform comparably to a model-free baseline that predicts by averaging the last three observations. We investigate whether their existing multi-step forecasts contain information discarded by conventional scoring. We propose Forecast Disagreement Anomaly Scoring (FDAS), a post-hoc scoring layer that combines the one-step forecasting error with disagreement among forecasts for the same target time made at different forecast origins. FDAS smooths and normalises both signals using a reference segment presumed to contain normal observations, then takes their maximum as the anomaly score. FDAS reuses forecasts already produced by a TSFM, without fine-tuning it or using anomaly labels for the target series. On the full TSB-AD benchmarks, we evaluate FDAS on eight univariate and five multivariate TSFM configurations. Compared with conventional scoring, FDAS raises mean VUS-PR from 0.29–0.30 to 0.45–0.50 on the univariate benchmark and from 0.21–0.24 to 0.37–0.40 on the multivariate benchmark. These gains are statistically significant in every evaluated configuration. Ablations show that combining the smoothed forecasting-error and disagreement tracks outperforms either track alone. FDAS makes frozen forecasting TSFMs competitive with dedicated anomaly-detection TSFMs, indicating that their limitation lies largely in the scoring rule rather than the forecasts.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series anomaly detection, time-series foundation model

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

This paper proposes FDAS, a forecasting-based anomaly detection approach that uses disagreement among forecasts for the same target time obtained from different forecast origins. Combined with smoothing and normalization, the proposed scoring method improves anomaly detection performance over conventional forecast-based scoring in the reported experiments.

Critical strengths
  1. Using forecasts from different origins: Comparing forecasts for the same target time from different origins is an interesting idea. Using frozen TSFMs, the paper shows that a post-hoc scoring method can improve anomaly detection performance.
  2. Potential of forecasting-based TSFMs for anomaly detection: The paper demonstrates that forecasting-based TSFMs can be used for anomaly detection by changing only the scoring method. This may broaden the applicability of existing forecasting-based TSFMs.
Critical weaknesses
  1. Contribution of forecast disagreement: Table 3 shows that EWMA smoothing accounts for much of the performance improvement over conventional scoring, particularly on the univariate benchmark. Adding forecast disagreement provides only modest additional gains, which makes the contribution of disagreement itself less convincing.
  2. Practical advantage over AD TSFMs: Although the proposed approach demonstrates the potential of forecasting-based TSFMs for anomaly detection, the referee is not convinced that they offer a practical advantage over TSFMs specifically designed for anomaly detection.
Questions for authors
  1. Beyond the aggregate results in Table 3 and the significance tests in Appendix E.1, how broadly are the additional gains from disagreement distributed across the evaluation series? How often does adding disagreement improve or degrade performance relative to smoothed and normalized forecast-error scoring? Are the improvements concentrated in particular types of anomalies?
  2. Under what conditions should a forecasting-based TSFM with FDAS be preferred to an anomaly-detection TSFM? Please distinguish between reusing an existing forecaster and deploying one specifically for anomaly detection, and clarify the trade-off between detection performance and computational cost in these settings.
3 comments
AuthorsOpenReview ↗

Response to Reviewer TvjS (1/3)

Thank you for finding the idea of comparing forecasts made at different origins for the same time step interesting, and for noting that changing only the scoring broadens the use of existing forecasting TSFMs. Your careful review and specific questions helped us sharpen the paper. Revised text is marked in blue in the revised PDF, and section, table and figure numbers below refer to it. We answer each point below.

TvjS-W1 (Contribution of forecast disagreement)

Weakness 1: "Contribution of forecast disagreement: Table 3 shows that EWMA smoothing accounts for much of the performance improvement over conventional scoring, particularly on the univariate benchmark. Adding forecast disagreement provides only modest additional gains, which makes the contribution of disagreement itself less convincing."

Disagreement improves mean VUS-PR beyond the smoothed and normalised one-step forecasting-error track (Ẽ1) in all 13 configurations, by 0.017–0.028 on TSB-AD-U (U) and 0.025–0.045 on TSB-AD-M (M), significantly in eight under unadjusted paired Wilcoxon tests (Appendix E.1). Smoothing and normalisation are components of the scoring layer together with the disagreement track, and the gain over conventional scoring (CS) comes from the layer as a whole. The submission did not report each component's contribution as a number; the revised abstract and main text now do.

Because the gains of smoothing and disagreement are not additive, the increment depends on whether smoothing is applied first. Without smoothing and with the same per-channel normalisation and channel maximum, the disagreement track alone outperforms the one-step forecasting-error track alone in 12 of 13 configurations (7/8 U, 5/5 M), and adding it raises mean VUS-PR by 0.07–0.10 on U and 0.03–0.07 on M (Table 3). Smoothing raises the one-step error track much more than the disagreement track (on U, +0.13–0.18 against −0.01 to +0.02), which is why the increment after smoothing is smaller.

Q1 below gives how this gain is distributed across series and anomaly types.

Revision: the abstract states the gain over conventional scoring with the same smoothing; the contributions in Section 1 and Section 5.2 state the disagreement increment with and without smoothing.

AuthorsOpenReview ↗

Response to Reviewer TvjS (2/3)

This comment continues our response (2/3).

TvjS-Q1 (Distribution of the gains from disagreement)

Question 1: "Beyond the aggregate results in Table 3 and the significance tests in Appendix E.1, how broadly are the additional gains from disagreement distributed across the evaluation series? How often does adding disagreement improve or degrade performance relative to smoothed and normalized forecast-error scoring? Are the improvements concentrated in particular types of anomalies?"

In all 13 configurations, the gain from the disagreement track is positive on average for series that contain only sequence anomalies, which make up 86% of U and 93% of M series, and for the half of the series with the longest average anomaly lengths. We re-aggregated the stored per-series results as a post-hoc analysis, using the baseline in your question, the smoothed and normalised forecasting-error track (Ẽ1). The values below are the change in VUS-PR when the disagreement track is combined with Ẽ1 (FDAS − Ẽ1), averaged across configurations.

Share of series improved or worsened (Table 4, "All series" rows). Compare the bold Improved and Worsened columns. Ties arise mainly in series scored through the fallback, where FDAS and Ẽ1 coincide (Appendix C.3). Percentages may not sum to 100 because of rounding.

ImprovedTiedWorsened
U (8 configurations)37%25%39%
M (5 configurations)55%10%35%

By anomaly type (point and sequence anomalies as labelled in TSB-AD; Table 4). Bold values show the mean VUS-PR change; parentheses give the number of configurations with a positive group mean (Appendix E.4).

Series (U / M)U: FDAS − Ẽ1M: FDAS − Ẽ1
Only sequence anomalies301 / 167+0.026 (8/8)+0.041 (5/5)
Containing point anomalies49 / 13−0.005 (0/8)−0.050 (0/5)

By quartile of average anomaly length, Q1 the shortest and Q4 the longest (Table 21, Figure 17). Each cell gives the number of series and the mean VUS-PR change, FDAS − Ẽ1, in bold.

QuartileQ1Q2Q3Q4
U92 / −0.01583 / +0.00688 / +0.06387 / +0.035
M45 / −0.05045 / +0.10245 / +0.05345 / +0.031

By type, the group mean is positive in every configuration for series that contain only sequence anomalies (U 8/8, M 5/5) and in no configuration for series that contain point anomalies. Length shows the same pattern: in all 13 configurations, the group mean is negative in Q1 and positive in Q3 and Q4, which contain half of the series on each benchmark. In Q2, it is positive in five of the eight U configurations and in all five M configurations. Anomalies are shorter on U, with a median average anomaly length of 116 time steps against 241 on M, which corresponds to the smaller overall gain on U.

Separately, comparing the two single tracks, the difference VUS-PR(D̃) − VUS-PR(Ẽ1) is positively correlated with the average anomaly length in all 13 configurations (Spearman; Figure 16).

Revision: Section 5.2 (Table 4) summarises the distribution, and Appendix E.4 (Table 21, Figures 16–17) gives the full results by anomaly type and length.

AuthorsOpenReview ↗

Response to Reviewer TvjS (3/3)

This comment continues our response (3/3).

TvjS-W2 + TvjS-Q2 (Practical advantage over AD TSFMs; conditions and cost)

Weakness 2: "Practical advantage over AD TSFMs: Although the proposed approach demonstrates the potential of forecasting-based TSFMs for anomaly detection, the referee is not convinced that they offer a practical advantage over TSFMs specifically designed for anomaly detection."

Question 2: "Under what conditions should a forecasting-based TSFM with FDAS be preferred to an anomaly-detection TSFM? Please distinguish between reusing an existing forecaster and deploying one specifically for anomaly detection, and clarify the trade-off between detection performance and computational cost in these settings."

We answer the two points together because they are connected. Which of FDAS and an AD TSFM is more suitable depends on the deployment setting:

  • If a forecaster already produces, at every time step, the H-step forecasts that FDAS needs, FDAS provides anomaly scores from those forecasts with no additional forward pass.
  • If a model is deployed on univariate data solely for anomaly detection, TimeRCD has higher mean VUS-PR and is faster.
  • On multivariate data, three FDAS configurations have higher mean VUS-PR than every AD TSFM compared in Table 2, and they processed all series of the evaluation split on one 32 GB GPU (RTX 5090).
  • With their released implementations and settings, TimeRCD and TSPulse exceeded the memory of the same GPU on some multivariate series.

The times in 1) and 3) below were measured over the full evaluation split (U: 350, M: 180 series), with forecasting and AD TSFMs on an RTX 5090 GPU and FDAS scoring on one CPU thread.

1) Reusing an existing forecaster. If the forecaster already produces the H-step forecasts that FDAS needs, FDAS uses them as they are, and the only added cost is scoring: 10.5–21.5 s on U and 3.3–6.6 min on M over the full evaluation split (Appendix A.2, Table 5). For the same series, the scoring cost grows linearly with H.

2) An existing forecaster that outputs a shorter horizon. For the nine backbones whose output length can be set, increasing the output length from H = 1 to the selected H changed the forecasting time by a factor of 0.98–1.14, measured on the same ten evaluation series per benchmark (Table 6).

3) Deploying a forecaster for anomaly detection. Forecasting dominates the cost, and the total time ranges from 1.2 min to 31.2 h on U and from 13.3 min to 35.0 h on M depending on the backbone. On U, TimeRCD (0.95 min) has higher VUS-PR and is faster than every FDAS configuration. On M, of the three configurations with higher mean VUS-PR than the AD TSFMs, the two with the light TTM-r2.1 backbone take 13.3 min and 17.1 min in total (Table 5). CHARM's TSB-AD implementation obtains embeddings from a hosted API, so its time could not be measured on the same hardware and is not in the table.

4) Trade-off between performance and cost. The number of forecasts for a time step equals H, so we recomputed FDAS with the first H steps of the stored forecasts, from H = 2 up to the value selected on the tuning split. FDAS − Ẽ1 is 0.003–0.014 on U and 0.000–0.005 on M at H = 2, and 0.017–0.028 and 0.025–0.045 at the selected H, larger in all 13 configurations, although it does not increase at every step of H (Appendix E.4, Table 22, Figure 18). When a forecaster is reused, the backbone is fixed and the added cost is scoring, which grows linearly with H. In a new deployment, forecasting accounts for most of the total time; across the nine evaluated backbones, a larger H changed the forecasting time by a factor of only 0.98–1.14, whereas the total time in Table 5 varied widely across backbones. A new deployment should therefore weigh the total time of each backbone together with its detection performance.

Revision: the contributions in Section 1 and Section 5.2 state the reuse cost and the comparison for a new deployment; the measurements and conditions for 1)–3) are in Appendix A.2 (Tables 5–6), and the results for 4) are in Appendix E.4 (Table 22, Figure 18).

We hope these answers address your concerns. If anything remains unclear, we would be glad to clarify.

Reviewer_4nLfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces an anomaly scoring scheme called Forecast Disagreement Anomaly Scoring (FDAS), which leverages a foundation model to generate forecasts and identifies anomalies based on the disagreement between the predicted values and the ground truth. FDAS proposes two tracks: one-step forecasting disagreement and forecasting errors over a temporal window. Each track is smoothed and normalized before aggregation into the final anomaly score.

Critical strengths

• A variety of pretrained models and evaluation metrics are used to provide a fair and comprehensive comparison, with additional comprehensive evaluations provided in the supplementary material. • Leverages existing pretrained time series models for time-series anomaly detection (TSAD).

Critical weaknesses

• Using forecasting errors for TSAD is not new. Therefore, the manuscript should provide a clearer explanation of the novelty of Forecast Disagreement Anomaly Scoring (FDAS). For example, L117–L118 could be elaborated to explain that “Both methods thus rely on forecasting-error scores, whereas our work additionally uses forecast disagreement.” Alternatively, if the novelty lies in the proposed track or channel aggregation strategy, this should be highlighted more clearly, and different aggregation methods should be evaluated in the ablation study.

Questions for authors

• Given that the novelty of this work lies in the proposed FDAS, it would be beneficial to evaluate FDAS using existing forecasting models instead of pretrained models to determine whether this approach can achieve improved performance. • A stronger justification of why FDAS is novelty as compare existing method are required.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4nLf (1/3)

Thank you for noting that the comparison is fair and comprehensive across a variety of pretrained models and evaluation metrics, and that the work leverages existing pretrained time series models for time-series anomaly detection. Revised text is marked in blue in the revised PDF, and section, table and figure numbers below refer to it. We answer each point below.

4nLf-S (Paper Summary: definition of the two tracks)

Paper Summary: "… identifies anomalies based on the disagreement between the predicted values and the ground truth. FDAS proposes two tracks: one-step forecasting disagreement and forecasting errors over a temporal window. Each track is smoothed and normalized before aggregation into the final anomaly score."

Forecast disagreement measures the differences among forecasts for the same target time step, whereas the forecasting error measures the difference between a forecast and the observation. The description that each track is smoothed and normalised before they are combined matches the paper.

  • Track 1, forecast disagreement (Section 4.2, Eq. (2)). The variance of the H forecasts that one frozen TSFM makes for the same target time step t from the forecast origins t−1, …, t−H. The forecasts are compared with each other, not with the ground truth, so this track can be computed before y_t is observed.
  • Track 2, one-step forecasting error (Section 4.2, Eq. (3)). The squared error between the one-step forecast and the observation y_t, the same quantity as in conventional scoring. This track measures the difference from the ground truth.
  • Combination (Sections 4.3–4.4, Eqs. (4)–(8)). Each track is smoothed with an EWMA (λ=0.95) and normalised by the 0.99-quantile of the train segment, and their maximum is taken. For multivariate series, the maximum over the channel scores is used.
  • Why both tracks are used. The forecasts can agree with each other while all deviating from y_t, in which case disagreement stays low; the forecasting-error track covers this case. Conversely, for an observation of 0 and forecasts (1, 1, 1) and (1, 1, 4) ordered by lead, both have one-step squared error 1, but their disagreement is 0 and 2.

Revision: to make this distinction clearer, we elaborated the sentence in Section 2 that you pointed out (L117–118 of the submission): "Our work additionally uses forecast disagreement, quantified as the variance of forecasts for the same target time step made at different forecast origins."

4nLf-W1a + 4nLf-Q2 (Novelty of FDAS)

Weakness 1: "Using forecasting errors for TSAD is not new. Therefore, the manuscript should provide a clearer explanation of the novelty of Forecast Disagreement Anomaly Scoring (FDAS). …"

Question 2: "A stronger justification of why FDAS is novelty as compare existing method are required."

We answer the two points together. As you note, using forecasting errors as anomaly scores is not new; the paper treats it as conventional scoring (CS) and uses it as the baseline. What FDAS adds is a scoring layer that obtains forecast disagreement, a signal different from the forecasting error, from a single frozen TSFM and combines it with the one-step forecasting error; with matched processing, adding this signal raises mean VUS-PR in all 13 configurations. LSTM-AD and W1-ACAS score the forecasting errors of forecasts from several origins (LSTM-AD fits a Gaussian to the error vector; W1-ACAS calibrates the individual errors), and FATE obtains disagreement from an ensemble of forecasting models. FDAS scores the dispersion of one TSFM's existing forecasts, without additional models or backbone inference.

Evidence.

  • Removing the disagreement track from FDAS lowers mean VUS-PR in all 13 configurations, by 0.017–0.028 on TSB-AD-U and 0.025–0.045 on TSB-AD-M, significantly in eight under unadjusted paired Wilcoxon tests (Appendix E.1).
  • Without smoothing and with the same normalisation and channel maximum, the disagreement track alone outperforms the one-step error track alone in 12 of 13 configurations (Table 3).
  • The VUS-PR difference between the two tracks, VUS-PR(D̃) − VUS-PR(Ẽ1), is positively correlated with the average anomaly length of a series in all 13 configurations (Spearman 0.44–0.53 on TSB-AD-U, 0.13–0.47 on TSB-AD-M; Figure 16).
  • Appendix E.4 reports the gain from adding the disagreement track (FDAS − Ẽ1) by anomaly length, anomaly type, and number of forecasts per target time step.
  • The scoring layer as a whole raises mean VUS-PR over CS by 0.15–0.21, significantly in all 13 configurations.

Revision: Section 2 now defines disagreement as above. The abstract, the contributions in Section 1 and Section 5.2 state the disagreement track's contribution; Appendix E.4 (Figures 16–17, Table 21) reports when its gain arises; Appendix A.4 (Figure 5) separates the components used by conventional scoring from those that FDAS adds.

AuthorsOpenReview ↗

Response to Reviewer 4nLf (2/3)

This comment continues our response (2/3).

4nLf-W1b (Comparison of aggregation methods)

Weakness 1: "Alternatively, if the novelty lies in the proposed track or channel aggregation strategy, this should be highlighted more clearly, and different aggregation methods should be evaluated in the ablation study."

Appendices E.2 and E.3 compare aggregation methods, and we do not claim novelty for the aggregation rules: mean and maximum aggregation differ by at most 0.01 in every configuration, whereas the minimum and the channel median score lower.

  • Track aggregation (Appendix E.2, Table 19; Table 15 of the submission). The maximum of Eq. (6) is compared with the arithmetic mean, product and minimum of the same two normalised tracks. Averaged over configurations, the mean is within 0.003 of the maximum on each benchmark, the minimum is the lowest in every configuration, and the product lies between the minimum and the mean.
  • Channel aggregation (Appendix E.3, Table 20; Table 16 of the submission). The channel maximum of Eq. (8) is compared with the channel mean and median. The channel mean is on average 0.007 above the maximum, and the median is the lowest in all five configurations.

In both cases the mean and the maximum perform similarly. We use the maximum because it preserves a high anomaly score that appears in only one track or in only some channels (Sections 4.3 and 4.4). The contribution of the paper is the scoring layer as a whole rather than an aggregation rule, as summarised in the novelty answer above.

Revision: none; the comparisons are in Appendices E.2 and E.3.

AuthorsOpenReview ↗

Response to Reviewer 4nLf (3/3)

This comment continues our response (3/3).

4nLf-Q1 (Evaluation with forecasting models that are not pretrained)

Question 1: "Given that the novelty of this work lies in the proposed FDAS, it would be beneficial to evaluate FDAS using existing forecasting models instead of pretrained models to determine whether this approach can achieve improved performance."

With forecasts from trained forecasting models, FDAS also outperforms conventional scoring (CS). Instead of pretrained TSFMs, we evaluated eight configurations of models trained on each series (four on TSB-AD-U, four on TSB-AD-M). On the same forecasts, FDAS exceeds CS in all eight, raising VUS-PR by 0.05–0.08 on TSB-AD-U and 0.06–0.14 on TSB-AD-M.

Setup. We used widely used forecasting models with the default settings of their public implementations and (L,H) = (96,96) for all. For each evaluation series, a separate model was trained on the first 80% of its train segment, and the remaining 20% was used for validation and early stopping. No tuning split or hyperparameter search was used, and the scoring layer and its parameters (λ, a) are those of the paper. For TSB-AD-M, we used only models designed to model dependencies between channels.

TSB-AD-U (350 series). Each cell shows CS → FDAS. The trained-forecaster rows are not in the paper; the last row gives the ranges over the frozen TSFMs in Tables 1–2. Bold highlights VUS-PR, the primary metric.

ForecasterVUS-PRVUS-ROCR-based-F1AUC-PRAUC-ROCStandard-F1
DLinear0.35 → 0.410.81 → 0.810.32 → 0.420.32 → 0.350.73 → 0.740.37 → 0.40
TSMixer0.37 → 0.450.83 → 0.850.37 → 0.450.35 → 0.380.75 → 0.790.40 → 0.44
Informer0.39 → 0.450.83 → 0.850.39 → 0.460.38 → 0.400.77 → 0.790.44 → 0.45
Autoformer0.31 → 0.360.76 → 0.780.33 → 0.370.29 → 0.290.68 → 0.700.33 → 0.35
Frozen TSFMs in the paper0.29–0.30 → 0.45–0.500.74–0.75 → 0.80–0.830.34–0.35 → 0.41–0.450.28–0.29 → 0.34–0.390.67–0.69 → 0.73–0.770.33–0.35 → 0.39–0.44

TSB-AD-M (180 series).

ForecasterVUS-PRVUS-ROCR-based-F1AUC-PRAUC-ROCStandard-F1
iTransformer0.27 → 0.410.69 → 0.770.29 → 0.340.26 → 0.350.64 → 0.720.31 → 0.42
SOFTS0.28 → 0.410.69 → 0.770.29 → 0.340.27 → 0.350.64 → 0.720.31 → 0.42
TimeXer0.29 → 0.410.70 → 0.780.30 → 0.340.28 → 0.340.65 → 0.730.32 → 0.42
TSMixer0.33 → 0.390.76 → 0.780.31 → 0.420.33 → 0.330.72 → 0.740.38 → 0.41
Frozen TSFMs in the paper0.21–0.24 → 0.37–0.400.67–0.70 → 0.74–0.760.25 → 0.28–0.350.20–0.23 → 0.32–0.350.62–0.65 → 0.69–0.720.24–0.27 → 0.37–0.41

What these results show. FDAS does not assume pretraining. The two assumptions in Section 4.1 are that one forecast sequence is used per lookback window (A1) and that a single forward pass outputs H ≥ 2 steps (A2); all eight models satisfy them. The gain appears across linear (DLinear), MLP (TSMixer), time-step-token Transformer (Informer, Autoformer) and channel-dependency architectures (iTransformer, SOFTS, TimeXer), so it is not tied to a particular forecaster. This is consistent with the paper's claim, which concerns the scoring rule rather than a particular forecaster. These runs evaluate the full pipeline; our analysis of the incremental benefit of disagreement concerns the evaluated frozen TSFMs.

Why the paper targets frozen TSFMs. The two settings rest on different deployment assumptions. A trained forecaster learns all its parameters on the train segment of each target series, so every series needs its own training (1,400 models on TSB-AD-U and 720 on TSB-AD-M in this experiment). A frozen TSFM applies one model to unseen series without parameter updates. We target this setting for two reasons:

  • First, forecasting TSFMs forecast unseen series without training, so wherever a TSFM is already used for forecasting, FDAS obtains anomaly scores from the same forecasts without additional models or training.
  • Second, under CS, forecasting TSFMs have been reported to perform comparably to the model-free Last-3 baseline (Zhu et al., 2026); with a frozen backbone the forecasts are fixed, so changes in detection performance can be attributed to the scoring layer.

Where per-series training is feasible, trained forecasters are also valid backbones for FDAS.

Revision: Section 4.1 now states that its assumptions do not require pre-training. These forecasters lie outside the frozen TSFMs that the paper studies, so the results are given in this response; we would be glad to add them to an appendix if you find them useful.

We hope these answers address your concerns. If anything remains unclear, we would be glad to clarify.

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

Time series foundation models (TSFMs) are pre-trained on large-scale time series datasets and can support downstream tasks such as anomaly detection, either directly or indirectly through forecasting. In the indirect case, the literature mainly scores anomalies with the one-step forecast error, which performs no better than simple baselines. The authors propose Forecast Disagreement Anomaly Scoring (FDAS), a post-hoc scoring layer for frozen forecasting TSFMs. It combines the one-step forecast error with the disagreement among forecasts made for the same time step from different forecast origins. Both signals are smoothed, normalised on the train segment, and combined by taking their maximum. No training or anomaly labels are needed for the target series. On the full TSB-AD-U and TSB-AD-M benchmarks (13 backbone configurations), FDAS raises mean VUS-PR from about 0.30 to 0.45–0.50 (univariate) and from about 0.22 to 0.37–0.40 (multivariate) compared to one-step error scoring. According to the authors, this makes forecasting TSFMs competitive with dedicated anomaly detection TSFMs.

Critical strengths

Strength 1:

  • A simple yet effective approach. FDAS acts as a post-hoc layer that can be applied without retraining the backbone TSFM or fitting additional parameters.

Strength 2:

  • The paper presents a concise and reproducible evaluation on the TSB-AD-U and TSB-AD-M benchmarks. The experimental study covers 13 configurations and includes per-series results, significance testing, sensitivity analyses, and ablation studies.
Critical weaknesses

Weakness 1:

  • Based on the ablation results in Table 3, the majority of the performance gains appear to stem from the normalization and smoothing of the one-step forecasting error, rather than from the proposed disagreement component that constitutes the main novelty of FDAS.
Questions for authors

Clarification requests:

  1. The abstract and contributions report the gain relative to unsmoothed conventional scoring. Table 3 shows that smoothed one-step error forecasting (CS^EWMA) already reaches 0.43–0.47 on univariate series, and Table 8 shows the added gain of FDAS over Ẽ1 is only about 0.02–0.03. Could the authors state the gain of FDAS over CS^EWMA in the abstract and the main text, and clarify which share of the claimed improvement is due to the disagreement track?
  2. (L, H) is selected to maximize VUS-PR under FDAS on the tuning splits, and conventional scoring then reuses the same pair. The tuning splits also share datasets with the evaluation splits. Could the authors tune (L, H) separately for conventional scoring to check that the comparison and the “label-free” claim are not favourably biased?
4 comments

As no rebuttal has been provided by the authors, I am keeping my review unchanged.

We sincerely apologise for the delay in our rebuttal. We have completed the additional analyses and are currently finalising our written response. We would greatly appreciate your patience for a little longer, and we apologise again for the delay.

AuthorsOpenReview ↗

Response to Reviewer oKzv (1/2)

Thank you for viewing FDAS as a simple and effective post-hoc layer that needs no backbone retraining or additional parameters, and for describing the evaluation, with per-series results over 13 configurations, significance tests, and sensitivity and ablation analyses, as concise and reproducible. Your specific questions helped us state the results more clearly. We added the requested numbers to the revised paper and re-ran conventional scoring with separately tuned (L,H) in all 13 configurations. Revised text is marked in blue in the revised PDF, and section, table and figure numbers below refer to it.

oKzv-W1 + oKzv-Q1 (Source of the gains; gain over CS^EWMA)

Weakness 1: "Based on the ablation results in Table 3, the majority of the performance gains appear to stem from the normalization and smoothing of the one-step forecasting error, rather than from the proposed disagreement component that constitutes the main novelty of FDAS."

Question 1: "Could the authors state the gain of FDAS over CS^EWMA in the abstract and the main text, and clarify which share of the claimed improvement is due to the disagreement track?"

Without smoothing and under the same per-channel normalisation and channel maximum, the disagreement track alone (d̃) outperforms the one-step forecasting-error track alone (ẽ1) in 12 of 13 configurations (Table 3). On TSB-AD-U (U), where ẽ1 has the same VUS-PR as conventional scoring (CS), d̃ − ẽ1 ranges from −0.02 to +0.07 and is positive in seven of eight configurations; on TSB-AD-M (M), it is +0.03 to +0.09 in all five. Adding disagreement to ẽ1 raises mean VUS-PR by 0.07–0.10 on U and 0.03–0.07 on M. These differences use the rounded Table 3 scores.

Gain over CS^EWMA, now in the abstract and Section 5.2. With the same smoothing, FDAS exceeds CS^EWMA by 0.017–0.028 on U and 0.044–0.088 on M. On U, Ẽ1 and CS^EWMA have the same VUS-PR, so this entire difference is due to the disagreement track. On M, 0.025–0.045 of it (39–73%) is due to the disagreement track and the rest to the per-channel normalisation and channel maximum of the forecasting-error track. FDAS exceeds both CS^EWMA and Ẽ1 in mean VUS-PR in all 13 configurations.

Share of the claimed improvement. Of the full gain over CS (0.157–0.202 on U, 0.154–0.195 on M), the disagreement track accounts for 9–17% on U and 16–23% on M when it is added after smoothing, and for 38–60% and 20–38% when it is added before smoothing. The ablations show a smaller increment after smoothing in all 13 configurations. These are increments within an ablation order, not unique causal shares.

Disagreement is useful both as a standalone signal without smoothing and as a complementary signal after matched smoothing and normalisation.

Bold values show disagreement's gain after matched processing. Percentages give its share of FDAS − CS^EWMA. Values use unrounded Table 3 scores; the disagreement gain is also in Table 11 (Table 8 of the submission). Components are rounded separately.

TSB-AD-U (350 series). Normalisation adds zero, so FDAS − Ẽ1 = FDAS − CS^EWMA.

ConfigurationDisagreement: FDAS − Ẽ1FDAS − CS^EWMAFull gain: FDAS − CSSmoothing: CS^EWMA − CS
Time-MoE base+0.026 (100%)+0.026+0.202+0.176
Time-MoE large+0.017 (100%)+0.017+0.189+0.172
TimesFM 2.0+0.020 (100%)+0.020+0.182+0.163
TimesFM 1.0+0.017 (100%)+0.017+0.174+0.157
Chronos-Bolt base+0.025 (100%)+0.025+0.185+0.160
Chronos-Bolt small+0.018 (100%)+0.018+0.174+0.157
TTM-r2.1 L1+0.028 (100%)+0.028+0.164+0.136
TTM-r2.1 L2+0.025 (100%)+0.025+0.157+0.132
Average+0.022 (100%)+0.022+0.178+0.157

TSB-AD-M (180 series). The remaining gain over CS^EWMA comes from per-channel normalisation and channel maximum.

ConfigurationDisagreement: FDAS − Ẽ1FDAS − CS^EWMAFull gain: FDAS − CSSmoothing: CS^EWMA − CS
TiRex-2+0.045 (52%)+0.088+0.195+0.107
TTM-r2.1 L2+0.037 (54%)+0.069+0.160+0.091
TTM-r2.1 L1+0.025 (39%)+0.064+0.159+0.095
Chronos-2+0.030 (64%)+0.046+0.154+0.107
Toto 2.0 (22m)+0.033 (73%)+0.044+0.159+0.114
Average+0.034 (55%)+0.062+0.165+0.103

Appendix E.4 (Figure 17, Table 21) reports where the disagreement gain arises: in all 13 configurations, its mean is negative in the quartile of series with the shortest average anomaly lengths (Q1) and positive in the two longest quartiles (Q3 and Q4).

Revision: the abstract and Section 5.2 now state the gain over CS^EWMA and the disagreement track's contribution; the contributions in Section 1 state the disagreement increment with and without smoothing.

AuthorsOpenReview ↗

Response to Reviewer oKzv (2/2)

This comment continues our response (2/2).

oKzv-Q2 (Separate tuning of (L,H) for conventional scoring)

Question 2: "(L, H) is selected to maximize VUS-PR under FDAS on the tuning splits, and conventional scoring then reuses the same pair. The tuning splits also share datasets with the evaluation splits. Could the authors tune (L, H) separately for conventional scoring to check that the comparison and the “label-free” claim are not favourably biased?"

As requested, we tuned (L,H) separately for conventional scoring (CS) and re-ran the comparison, and we report CS^EWMA tuned in the same way. Separate tuning changed CS by at most 0.01 and CS^EWMA by at most 0.03, and FDAS still exceeds both in all 13 configurations.

1) Dataset sharing between the tuning and evaluation splits. This is a property of the official TSB-AD tuning and evaluation splits; we follow these splits and the evaluation protocol as released (Section 5.1, Appendix B.1), and the condition applies equally to all methods. In this experiment, CS and CS^EWMA select their own (L,H) on the same tuning split as FDAS, so all three methods choose their settings under the same condition. The two splits contain no common series (Appendix B.2).

2) Use of labels. Our claim is that FDAS uses no anomaly labels for the target series. As for other methods on TSB-AD, (L,H), λ and a are selected with the labels of the tuning split. We understand your concern as whether settings chosen with these labels fit FDAS alone and so favour it; the results below, in which CS receives the same tuning with the same labels, address this. Because the submitted sentence in Section 5.1, "Anomaly labels are used only to compute evaluation metrics, not for scoring.", did not state this use, we revised it to: "Anomaly labels are used to select (L, H), λ, and a on the tuning splits and to compute evaluation metrics, not for scoring."

3) Results with separately tuned (L,H). Using the same tuning series, the same (L,H) grid and the same selection criterion as FDAS (maximum mean VUS-PR on the tuning split), CS and CS^EWMA each select their own (L,H); we then re-evaluated on the full evaluation split (U: 350, M: 180 series). CS^EWMA uses λ=0.95 as in FDAS, and FDAS keeps the values in the paper. Each cell shows the selected (L,H) and the mean VUS-PR; p is the unadjusted paired Wilcoxon p-value of FDAS against independently tuned CS^EWMA (Table 18, Appendix D.7).

Bold scores mark the highest mean VUS-PR in each row after separate tuning.

ConfigurationFDASCSCS^EWMAp
UTime-MoE (base)(512,32); 0.50(256,32); 0.30(256,32); 0.460.004
UTime-MoE (large)(512,64); 0.49(64,64); 0.30(256,64); 0.470.072
UTimesFM 2.0(512,32); 0.48(256,32); 0.30(256,32); 0.45<0.001
UTimesFM 1.0(512,64); 0.47(256,64); 0.30(256,32); 0.450.064
UChronos-Bolt (base)(512,32); 0.48(256,64); 0.29(128,32); 0.440.004
UChronos-Bolt (small)(512,32); 0.46(256,64); 0.29(64,64); 0.450.12
UTTM-r2.1 (L1)(512,48); 0.46(90,30); 0.29(512,48); 0.430.017
UTTM-r2.1 (L2)(512,48); 0.45(90,30); 0.29(512,48); 0.430.16
MTiRex-2(64,64); 0.40(512,32); 0.22(512,64); 0.340.004
MTTM-r2.1 (L2)(512,48); 0.40(512,48); 0.24(52,16); 0.31<0.001
MTTM-r2.1 (L1)(512,96); 0.40(180,60); 0.23(360,60); 0.32<0.001
MChronos-2(256,64); 0.38(512,64); 0.23(512,64); 0.340.004
MToto 2.0 (22m)(256,64); 0.37(512,64); 0.22(512,64); 0.340.11
  • FDAS exceeds independently tuned CS by 0.16–0.20 on U and 0.15–0.18 on M, significantly in all 13 configurations.
  • FDAS exceeds independently tuned CS^EWMA in all 13 configurations, significantly in eight (four U, four M) under unadjusted paired Wilcoxon tests.

Revision: the results are in the new Appendix D.7 (Table 18) and summarised in Section 5.2, Appendix B.4 points to D.7 after the description of the (L,H) selection, and Section 5.1 states that the labels of the tuning split select (L,H), λ and a.

We hope these answers address your concerns. If anything remains unclear, we would be glad to clarify.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.