acceptodds
Under review as a conference paper at ICLR 2027

QPath: Joint Anomaly Detection from Marginal Quantile Forecasts

Abstract

Time-series foundation models increasingly use direct quantile forecasting to provide efficient uncertainty estimates without specifying a parametric predictive distribution. Yet marginal quantiles leave dependence across horizon steps unspecified: a path can be ordinary at individual steps while being jointly anomalous. QPath addresses this gap at the detection stage without changing the forecaster or constructing a joint predictive distribution. It combines the mean quantile loss across forecast quantiles with a temporal direction factor estimated from historical median residuals, followed by reference normalization and channel averaging. We evaluate QPath on synthetic anomaly patterns and the univariate and multivariate TSB-AD benchmarks using Chronos-2 and TimesFM 2.5. Across zero-shot and head-only fine-tuned settings, average relative improvements in whole-series VUS-PR over a quantile-loss-only baseline (quantile error) are 22.1% on TSB-AD-U and 4.3% on TSB-AD-M; the univariate absolute gains are 0.069–0.074. Multivariate gains remain inconclusive under source-cluster inference. Removing explicit cross terms eliminates the mean univariate gain in a matched diagonal control; replacing absolute error with quantile loss adds 0.003–0.010 on TSB-AD-U.

Primary area:
learning on time series and dynamical systems
Keywords:
Timeseries anomaly detection, Quantile Forecasting, Forecasting based anomaly detection, Time Series Foundation Model

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_HfybOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes QPath a method to evaluate how different a quantile forecast is from the ground truth. QPath uses the Rayleigh quotient of the median residuals and its precision matrix to score how anomalous a quantile forecast is. The authors provide many experiments and ablations on the TSB-AD dataset with Chronos 2 and TimesFM time series foundation models. The results show that QPath performs better than the quantile error for anomaly detection. The paper also looks at some ablations of QPath, including only using the diagonal entries of the precision and using Mahalanobis distance instead of QPath.

Critical strengths
  • Figure 3 is nice and shows both what the task is and the behaviour of QPath

  • The authors provided statistical testing for some of their results, even though it is unsurprising given the sample size of significance.

Critical weaknesses
  • The paper is quite hard to read. Many paragraphs do not contain a starting/topic sentence and many non sequiturs. This needs to be fixed for me to adjust my score.

  • The paper does not state that QPath is the Rayleigh quotient of the median residuals and its precision matrix. Nor that Rayleigh quotients have been extensively used in ML, including for anomaly detection. This significantly limits the novelty of the paper and reference to the Rayleigh quotient and its use in ML should be given.

  • The paper does not compare to existing anomaly detection methods not calibration-based methods, only to quantile error in the main results. Therefore, it is hard to understand the performance of QPath with respect to existing work.

  • As a minor point, I think how VUS-PR works should be explained in the text somewhere, maybe the appendix.

Questions for authors
  • Why for variance burst does the QPath perform better? As I would naively think that the residuals would not be correlated in this case.
4 comments

General response

We sincerely thank the reviewer for the careful review and constructive feedback, and we apologize for posting our response late in the discussion period. We appreciate the positive remarks on Figure 3 (S1) and the statistical testing (S2). In this response and the revision, we address the readability of the text (W1), the Rayleigh quotient (W2), the comparison with existing methods (W3), the explanation of VUS-PR (W4), and the gain on variance bursts (Q1). Section and table numbers refer to the revised paper.

W1. Readability of the text

"The paper is quite hard to read. Many paragraphs do not contain a starting/topic sentence and many non sequiturs."

We sincerely apologize that the original manuscript was hard to read. In the revision, we (i) rewrote the main text so that every paragraph opens with the statement it develops, (ii) removed the non sequiturs by stating the reason that links each claim to the next and by removing the caveats that interrupted the argument, (iii) defined each term at first use, and (iv) reorganized the text, so that Section 3 opens with an overview and builds the score step by step with Figure 1, and the abstract and introduction are rewritten. For example, Sections 5.2 and 5.3 now open with their main result, Section 5.5 opens with what the two ablations test, and Section 5.1 now gives the reason why QPath helps on abrupt anomalies and not on trend changes.

W2. Naming the direction factor as a Rayleigh quotient

"The paper does not state that QPath is the Rayleigh quotient...This significantly limits the novelty of the paper..."

We thank the reviewer for pointing this out, and we apologize for not stating it. The squared direction factor of QPath is the Rayleigh quotient ρ(v)² = vᵀPv / vᵀv, where v holds the standardized residuals of a path and P is the inverse of the regularized covariance matrix of such residuals in the reference. The revision now states this (Sections 1, 2, and 3.3) and cites the statistical background of the quotient in Section 2, namely matrix analysis (Horn & Johnson, 2012), PCA, Fisher discriminant analysis, and the Mahalanobis distance, which is a standard score for anomaly and out-of-distribution detection (Lee et al., 2018).

The quotient itself is well known, and we exploit two of its properties. It depends only on the direction of v, not on the size of the error, and it is small for arrangements along which normal residuals vary together and large for arrangements that are rare. What is new is its use to model how the errors at different time steps relate to each other, which a quantile forecast, made for each time step separately, does not provide. The size of the error is measured by the quantile loss of the forecast, in place of the length of v that the Mahalanobis distance uses. In the tables, Full denotes evaluation on the whole series, as in the TSB-AD protocol, and Test on its test part alone.

Table R1. Full VUS-PR on U. "Diagonal of P only" removes the entries of P that couple different time steps.

ModelForecastQuantile errorDiagonal of P onlyMahalanobis formQPath
Chronos-2ZS0.3180.3160.3770.389
FT0.3260.3240.3820.395
TimesFM 2.5ZS0.3220.3190.3760.394
FT0.3260.3240.3760.400

As shown in Table R1, the gain disappears when the entries of P that couple different time steps are removed, so it comes from the relation between time steps. QPath also exceeds the Mahalanobis form, which applies the same quotient to the length of v, in all 16 settings (U and M, two forecasting models, ZS and FT, Full and Test) with p ≤ 0.011. Details are in Section 5.5 and Appendix H.2.

This comment continues our response (2/2).

W3. Comparison with existing methods in the main results

"The paper does not compare to existing anomaly detection methods not calibration-based methods, only to quantile error..."

Following the reviewer's suggestion, we have added to the main results a calibration-based method, W1-ACAS (Martinez Gil et al., ICLR 2026), and an existing detection score, the error-vector Gaussian (Malhotra et al., 2015). Both compute an anomaly score from the errors of the same quantile forecasts, as quantile error does. W1-ACAS calibrates the forecast errors and keeps adapting while it scores, and the error-vector Gaussian fits a Gaussian to the errors of the median forecast.

Table R2. VUS-PR on U, Full / Test (series and timestamps common to all four rules).

ModelForecastQuantile errorW1-ACAS [1]Error-vector Gaussian [2]QPath
Chronos-2ZS0.311 / 0.3240.203 / 0.2630.341 / 0.3580.383 / 0.397
FT0.319 / 0.3320.203 / 0.2620.343 / 0.3590.389 / 0.402
TimesFM 2.5ZS0.316 / 0.3300.211 / 0.2680.375 / 0.3910.389 / 0.401
FT0.320 / 0.3330.210 / 0.2670.376 / 0.3930.395 / 0.407

QPath is the highest in all eight settings of Table R2. Among the compared rules, it is the only one that uses how the errors at consecutive time steps of one forecast relate to each other. Among the published forecasting-based detectors of the TSB-AD leaderboard, QPath is also first, with 0.407 on U and 0.371 on M against 0.343 and 0.313. Details are in Sections 5.2 and 5.4 and Appendix H.1.

W4. Explanation of VUS-PR

"...how VUS-PR works should be explained in the text somewhere, maybe the appendix."

Following the reviewer's suggestion, Section 4.2 now explains VUS-PR with its formula. In short, each labeled anomaly range is extended by a buffer of ℓ time steps, the area under the precision–recall curve is computed for every ℓ from 0 to a window w, and VUS-PR is the average of these areas. It is therefore threshold-free and tolerant of a small time lag. Details are in Appendix B.4.

Q1. Why QPath improves variance bursts

"Why for variance burst does the QPath perform better? As I would naively think that the residuals would not be correlated in this case."

We appreciate this thoughtful question. QPath does not require the residuals of an anomaly to be correlated. It uses the correlation of normal residuals as a yardstick and raises a path whose arrangement is unusual under that correlation.

In our synthetic data, a variance burst amplifies the noise of the series, which is autocorrelated (AR(1) with coefficient 0.8), so the residuals inside a burst remain correlated as normal residuals are, and the direction factor does not single them out. From the definition of the score, the gain arises in two other places. First, at the start of a burst, a path that is calm and then abruptly large is a rare arrangement when neighboring residuals are positively correlated, so the direction factor raises it, as for a spike. Quantile error averages over the 64 steps of the path, so it is still small at this point. Second, the largest normal errors come from residuals that drift to one side, a common arrangement that the direction factor lowers, so fewer normal paths rank above the burst. We have added this analysis to Section 5.1.

We thank the reviewer again for the constructive comments, which have helped us improve the paper.

References

[1] Natalia Martinez Gil, Fearghal O'Donncha, Wesley M. Gifford, Nianjun Zhou, Dhaval C. Patel, and Roman Vaculin. Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring. ICLR 2026.

[2] Pankaj Malhotra, Lovekesh Vig, Gautam Shroff, and Puneet Agarwal. Long Short Term Memory Networks for Anomaly Detection in Time Series. ESANN 2015, pp. 89–94.

Reviewer_HfybOpenReview ↗

Reply to rebuttal

Dear Authors,

Thank you for your rebuttal that addressed many of my concerns. Given this I will increase my score and think the paper is much improved.

Thank you for your kind reply and positive assessment of our revision. Your feedback has greatly improved the paper.

If you have any remaining concerns, please let us know and we will do our best to address them.

Reviewer_GW1uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers time-series foundation models that use marginal quantiles to produce predictions. The core idea is that, the observation can appear to be normal at individual step levels but is anomalous as a whole. To address this, this paper proposes QPath, an anomaly detection post-hoc score that combines the mean quantile loss with a directional factor, that gauges how abnormal the prediction shape is. Experimental results show that QPath outperforms the quantile loss baseline on univariate time series

Critical strengths
  • Time-series foundation models are increasing using quantile heads, but they do not explicitly predict the joint distribution. Thus this paper considers a timely and important topic.
  • The gains on univariate data are inconsistent in the experimental result
Critical weaknesses
  • Section 3 is very dense and compact, and the formulas and variables are not sufficiently discussed
  • The main comparison is only against the standard quantile error
  • The gains on multivariate results, and the results against a Mahalanobis-type score are relatively small
Questions for authors
  • Are there other known baselines other than quantile error to compared with in this setting?

  • In the mulrivariate case, each channel is considered independently and then averaged. It appears that the core idea and the proposed approach can also be applied across channels. Specifically, each channel might appear normal but the relationship between channels can be anomalous. I wonder if the authors have tested on this

2 comments

General response

We sincerely thank the reviewer for the careful review and constructive feedback, and we apologize for posting our response late in the discussion period. We appreciate the positive remarks on the timeliness of quantile heads that do not predict the joint distribution (S1) and on the univariate gains (S2), which hold in all eight settings (0.068–0.074 in VUS-PR). In this response and the revision, we address the readability of Section 3 (W1), the comparison with baselines other than quantile error (W2, Q1), the size of the gains on multivariate data and over the Mahalanobis-type score (W3), and the relation between channels (Q2). Section and table numbers refer to the revised paper.

W1. Readability of Section 3

"Section 3 is very dense and compact, and the formulas and variables are not sufficiently discussed"

We sincerely apologize that Section 3 was dense and that its formulas and variables were not sufficiently explained. In the revision, we (i) open the section with an overview of the score as the size of the forecast error times a direction factor, (ii) state what each quantity measures and define every symbol at first use (Sections 3.1 and 3.2), and (iii) explain the direction factor as a Rayleigh quotient with explicit bounds, which lowers the score for arrangements of errors that are common in normal data and raises it for rare ones, with Figure 1 showing the geometry (Section 3.3).

W2, Q1. Comparison with baselines other than quantile error

"The main comparison is only against the standard quantile error...Are there other known baselines...?"

Following the reviewer's suggestion, we have added two published scoring rules to the main results, W1-ACAS (Martinez Gil et al., ICLR 2026) and the error-vector Gaussian (Malhotra et al., 2015). Both compute an anomaly score from the errors of the same quantile forecasts, as quantile error does. W1-ACAS calibrates the forecast errors and keeps adapting while it scores, and the error-vector Gaussian fits a Gaussian to the errors of the median forecast. In the tables, Full denotes evaluation on the whole series, as in the TSB-AD protocol, and Test on its test part alone.

Table R1. VUS-PR on U, Full / Test (series and timestamps common to all four rules).

ModelForecastQuantile errorW1-ACAS [1]Error-vector Gaussian [2]QPath
Chronos-2ZS0.311 / 0.3240.203 / 0.2630.341 / 0.3580.383 / 0.397
FT0.319 / 0.3320.203 / 0.2620.343 / 0.3590.389 / 0.402
TimesFM 2.5ZS0.316 / 0.3300.211 / 0.2680.375 / 0.3910.389 / 0.401
FT0.320 / 0.3330.210 / 0.2670.376 / 0.3930.395 / 0.407

QPath is the highest in all eight settings of Table R1. Among the compared rules, it is the only one that uses how the errors at consecutive time steps of one forecast relate to each other. Among the published forecasting-based detectors of the TSB-AD leaderboard, QPath is also first, with 0.407 on U and 0.371 on M against 0.343 and 0.313. Details are in Sections 5.2 and 5.4 and Appendix H.1.

References

[1] Natalia Martinez Gil, Fearghal O'Donncha, Wesley M. Gifford, Nianjun Zhou, Dhaval C. Patel, and Roman Vaculin. Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring. ICLR 2026.

[2] Pankaj Malhotra, Lovekesh Vig, Gautam Shroff, and Puneet Agarwal. Long Short Term Memory Networks for Anomaly Detection in Time Series. ESANN 2015, pp. 89–94.

This comment continues our response (2/2).

W3. Size of the gains on multivariate data and over a Mahalanobis-type score

"The gains on multivariate results, and the results against a Mahalanobis-type score are relatively small"

We thank the reviewer for this comment. Both gains are consistent across settings. On M, QPath has the higher mean in all 8 settings, with a score that is applied to each channel separately. Over the Mahalanobis form, QPath is higher in all 16 settings (U and M, two forecasting models, ZS and FT, Full and Test), with p ≤ 0.011 in each.

For the Mahalanobis-type score, we would like to clarify what the comparison measures. This score and QPath share the direction factor, which gives most of the gain over quantile error on U (Table R2), and differ only in the magnitude that the factor scales, the length of the standardized errors for the Mahalanobis form and the quantile loss for QPath. The comparison therefore isolates the choice of magnitude, and it shows that the quantile loss is the better magnitude in every setting.

Table R2. Full VUS-PR. p is the paired p value of QPath against the Mahalanobis form.

DataModelForecastQuantile errorMahalanobis formQPathp
UChronos-2ZS0.3180.3770.3890.010
FT0.3260.3820.3950.010
TimesFM 2.5ZS0.3220.3760.3940.001
FT0.3260.3760.4003e-5
MChronos-2ZS0.3480.3500.3670.006
FT0.3510.3520.3680.010
TimesFM 2.5ZS0.3440.3300.3536e-5
FT0.3470.3370.3623e-5

Against quantile error on M, the paired p value is below 0.05 in 4 of the 8 settings. QPath captures the relation between time steps within one channel, which may explain why its gain is larger on univariate data. Details are in Section 5.5 and Appendix H.2.

Q2. Applying the idea across channels

"...each channel might appear normal but the relationship between channels can be anomalous. I wonder if the authors have tested on this"

We appreciate this suggestion and tested it in the revision. QPath focuses on the relation between time steps that quantile forecasts leave out, and, as the reviewer suggested, we applied the same direction factor across channels.

Table R3. Full VUS-PR on M.

ModelForecastQuantile errorQPathQPath + channel factor
Chronos-2ZS0.3480.3670.342
FT0.3510.3680.343
TimesFM 2.5ZS0.3440.3530.325
FT0.3470.3620.338

As shown in Table R3, QPath remains the best of the three scores on M, and a direct transfer of the factor to channels did not improve on it. We expect that a factor designed for anomalous relations between channels will complement QPath, and we now state this direction in Section 6. Details are in Appendix H.4.

We thank the reviewer again for the constructive comments, which have helped us improve the paper.

Reviewer_6sttOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes QPath to address the missing information regarding temporal dependencies across horizon steps in time-series anomaly detection. The process neither alters the forecaster nor constructs a joint distribution for prediction. Essentially, the method calculates the average quantile loss across the forecast steps and integrates it with a temporal direction factor derived from past median residuals. The proposed method has been evaluated on synthetic anomaly patterns, two TSB-AD benchmarks, namely Chronos-2 and TimesFM 2.5 under settings such as Zero-Shot and Fine-Tuned utilizing VUS-PR (Volume Under the Surface of Precision-Recall) together with statistical paired t-test comparison.

Critical strengths

The paper addresses an important limitation of existing quantile-based anomaly detection methods, which mainly focus on prediction errors at individual time steps while overlooking temporal dependencies. The proposed method is simple, intuitive, and can be applied to existing forecasting models.

Critical weaknesses
  1. Discussion related to computational complexity and resource utilization is not found. It would be good to know about the computational aspects of the proposed method.

  2. The evaluation is limited to two forecasting models and one benchmark suite, although it includes both real-world and synthetic datasets. More extensive validation would strengthen the paper. In addition, the improvements on multivariate data are less convincing, and the method does not consistently improve the detection of trend-change anomalies.

Questions for authors

From the computational perspective, how the proposed method score compare with other competing methods would be useful. Moreover, determining the volume of training data needed for a reliable inference would be critical?

4 comments

General response

We sincerely thank the reviewer for the careful review and constructive feedback, and we apologize for posting our response late in the discussion period. We appreciate the positive remarks on the importance of temporal dependence for quantile-based detection (S1), the simplicity of the method (S2), and its applicability to existing forecasters (S3). In this response and the revision, we address the computational cost with measured runtimes (W1, Q1), the breadth of the evaluation with two published baselines (W2-1), the gains on multivariate data (W2-2), the trend-change anomalies (W2-3), and the amount of reference data needed (Q2). Section and table numbers refer to the revised paper.

W1, Q1. Computational cost and its comparison with other methods

"Discussion related to computational complexity and resource utilization is not found."

We thank the reviewer for pointing this out and have added this discussion to the paper. QPath is a lightweight post-processing step. Its cost grows linearly with the number of paths, as for quantile error, so the two differ only by a constant factor, and the only extra work is the solution of one 64 × 64 triangular system per path.

Table R1. Time to fit each scoring rule on n reference paths and to score 1,024 paths (one channel, one CPU thread).

Scoring rulen = 181n = 2,000
Quantile error7.67 ms20.17 ms
QPath9.75 ms26.74 ms
Error-vector Gaussian1.56 ms3.52 ms
W1-ACAS25.77 s69.56 s

As shown in Table R1, QPath takes about 1.3 times as long as quantile error. The table also includes the two published scoring rules that the revision adds, W1-ACAS (Martinez Gil et al., ICLR 2026) and the error-vector Gaussian (Malhotra et al., 2015). W1-ACAS calibrates the forecast errors and keeps adapting while it scores, and the error-vector Gaussian fits a Gaussian to the errors of the median forecast. QPath is more than 2,000 times as fast as W1-ACAS.

Regarding resource utilization, QPath has no trained parameters and works on forecasts that the model has already produced, so it needs no training and no additional forward pass of the forecaster, and its fitted model is small, one 64 × 64 matrix per channel. Details are in Section 3.3 and Appendix B.5.

W2-1. Number of forecasting models and benchmarks

"The evaluation is limited to two forecasting models and one benchmark suite..."

We agree with the reviewer that our real-data evaluation uses one benchmark suite, TSB-AD. This suite is itself a collection of public datasets. The part that we use contains 530 series, 350 univariate from 23 datasets and 180 multivariate from 17. By the labels of TSB-AD, they span nine domains (facility monitoring, web services, medicine, human activity, sensors, environment, finance, traffic, and synthetic data), and their training segments range from 500 to 50,000 time steps. Across this collection, QPath improves the mean VUS-PR in every setting, for both forecasting models, Chronos-2 and TimesFM 2.5, with zero-shot and fine-tuned forecasts. A synthetic study with five controlled anomaly types complements the benchmark. The main results of the revision also compare QPath with the two published scoring rules of Table R1 on the same forecasts. In the tables, Full denotes evaluation on the whole series, as in the TSB-AD protocol, and Test on its test part alone.

Table R2. VUS-PR on U, Full / Test (series and timestamps common to all four rules).

ModelForecastQuantile errorW1-ACAS [1]Error-vector Gaussian [2]QPath
Chronos-2ZS0.311 / 0.3240.203 / 0.2630.341 / 0.3580.383 / 0.397
FT0.319 / 0.3320.203 / 0.2620.343 / 0.3590.389 / 0.402
TimesFM 2.5ZS0.316 / 0.3300.211 / 0.2680.375 / 0.3910.389 / 0.401
FT0.320 / 0.3330.210 / 0.2670.376 / 0.3930.395 / 0.407

QPath is the highest in all eight settings of Table R2. Among the compared rules, it is the only one that uses how the errors at consecutive time steps of one forecast relate to each other. Among the published forecasting-based detectors of the TSB-AD leaderboard, QPath is also first, with 0.407 on U and 0.371 on M against 0.343 and 0.313. We agree that further forecasting models are a valuable extension, and QPath applies to any multi-horizon quantile forecaster without retraining it. Details are in Sections 4.1, 5.2, and 5.4 and Appendix H.1.

References

[1] Natalia Martinez Gil, Fearghal O'Donncha, Wesley M. Gifford, Nianjun Zhou, Dhaval C. Patel, and Roman Vaculin. Adaptive Conformal Anomaly Detection with Time Series Foundation Models for Signal Monitoring. ICLR 2026.

[2] Pankaj Malhotra, Lovekesh Vig, Gautam Shroff, and Puneet Agarwal. Long Short Term Memory Networks for Anomaly Detection in Time Series. ESANN 2015, pp. 89–94.

This comment continues our response (2/2).

W2-2. Size of the gain on multivariate data

"...the improvements on multivariate data are less convincing..."

We appreciate this comment. As shown in Table R3, QPath raises the mean VUS-PR on multivariate data in all 8 settings, with a score that is applied to each channel separately, without a model of the relation between channels.

Table R3. VUS-PR on M. p is the paired p value of QPath against quantile error.

ModelForecastFullTest
Quantile errorQPathpQuantile errorQPathp
Chronos-2ZS0.3480.3670.0060.3600.3740.032
FT0.3510.3680.0150.3630.3750.066
TimesFM 2.5ZS0.3440.3530.2650.3560.3600.504
FT0.3470.3620.0480.3580.3700.113

As noted under W2-1, QPath is also first among the published forecasting-based detectors on M (0.371 against 0.313). QPath captures the relation between time steps within one channel, which may explain why its gain is larger on univariate data, and the relation between channels is a direction for further gains.

W2-3. Trend changes, one of the five synthetic anomaly types

"...the method does not consistently improve the detection of trend-change anomalies."

We thank the reviewer for this observation. QPath improves four of the five anomaly types for both forecasting models, with the largest gains on spikes (VUS-PR from 0.061 to 0.264 with Chronos-2 and from 0.038 to 0.300 with TimesFM 2.5). On trend changes, the difference from quantile error is small and not statistically significant at the default anomaly strength.

The contrast between abrupt anomalies and trend changes follows from how the score works. The direction factor checks each error against the errors at the other time steps of the same path. In normal data, errors at neighboring time steps are positively correlated, so an abrupt departure such as a spike or the onset of a level shift is a rare arrangement, and QPath amplifies it. A trend change begins gradually and gives a smooth run of errors of one sign, which is common in normal data and is discounted, so it is detected through the quantile loss that QPath keeps. We have added this explanation of why QPath helps on abrupt anomalies and not on trend changes to Section 5.1.

Q2. Amount of reference data needed

"...the volume of training data needed for a reliable inference..."

We appreciate this question. QPath has no trained parameters and needs only the reference, the forecast errors on the label-free training segment of a series, and a few hundred reference paths give most of the gain. To measure this, we restricted the reference to n paths, which correspond to a training segment of n + 319 time steps (a context of 256 steps and the 64-step horizon).

Table R4. Gain of QPath over quantile error in Full VUS-PR when the reference is restricted to n paths (237 U series with at least 1,024 reference paths).

ModelForecastn=641281812565121,024all
Chronos-2ZS0.0050.0540.0610.0680.0780.0800.086
FT0.0060.0510.0590.0660.0770.0770.083
TimesFM 2.5ZS0.0050.0540.0560.0670.0780.0820.088
FT0.0050.0550.0560.0650.0770.0810.086

As shown in Table R4, the gain is significant from n = 128 (p < 1e-6), that is, from a training segment of 447 time steps, and it levels off as the reference grows. QPath therefore works with short training segments (n = 181 is the shortest reference in U). Details are in Appendix H.3.

We thank the reviewer again for the constructive comments, which have helped us improve the paper.

Reviewer_6sttOpenReview ↗

Assessment of the Authors' Response:

Assessment of the Authors' Response:

I appreciate the authors' detailed and constructive responses. The rebuttal provides substantial clarifications and additional experimental evidence that address several of my initial concerns. In particular, the computational analysis demonstrates that QPath introduces relatively modest overhead compared with standard quantile error scoring. The additional comparisons with W1-ACAS and the error-vector Gaussian baseline strengthen the experimental evaluation. Moreover, the reference-size study provides useful evidence that QPath can achieve meaningful improvements with relatively short reference histories. The diagonal ablation also supports the importance of temporal dependencies in anomaly detection.

Remaining Concerns:

  1. Although QPath consistently achieves higher average scores on multivariate datasets, the improvements remain relatively modest and dataset-dependent. The source-cluster confidence intervals include zero for all multivariate comparisons against quantile error, suggesting that broader generalization remains uncertain. Furthermore, the evaluation is still limited to two forecasting models and one benchmark suite. Additional validation would strengthen the conclusions.
  2. I acknowledge the authors' clarification regarding the Rayleigh quotient. However, the underlying mathematical formulation is well established, and the main contribution lies in its integration with quantile-based forecasting errors rather than a fundamentally new mathematical framework. Nevertheless, I consider this combination practically meaningful, particularly because it can be applied to existing forecasting models without additional training.
  3. Although the authors provide a reasonable explanation for the weaker performance on gradual trend changes, the synthetic experiments still indicate that QPath can underperform standard quantile error scoring in certain settings. Based on the above, I maintain my original rating.

We sincerely thank the reviewer for the careful assessment of our response.

(1) We agree that the multivariate gain is modest and that further models and benchmarks would strengthen the conclusions. QPath is still higher in all eight multivariate settings, with p < 0.05 in four by the paired test across series, and on univariate data even the source-cluster intervals are positive throughout.

(2) We agree that the Rayleigh quotient is well established, and the name gives a concise description of one part of QPath, its direction factor. QPath uses this known quantity to capture the dependence across time steps that quantile forecasts leave out, and, combined with the quantile loss.

(3) We agree that QPath is slightly below quantile error on gradual trend changes. This is expected in principle (Section 5.1), as the direction factor discounts smooth runs of errors and amplifies abrupt ones, and the other four anomaly types improve for both models.

We thank the reviewer again for the constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.