acceptodds
Under review as a conference paper at ICLR 2027

BandSAD: Band-Decoupled Forecasting with Whitening for Label-Free Time-Series Anomaly Detection

Abstract

Univariate time-series anomaly detection is dominated by deep autoencoding pipelines whose anomaly score has no interpretable scale, forcing practitioners to tune a decision threshold on labeled anomalies—a resource that streaming applications rarely have. We present BandSAD, a band-decoupled forecaster that (i) splits the forecasting target into frequency bands through a learnable partition-of-unity mask in the Fourier domain, (ii) whitens the resulting per-band errors with a robust MAD-based variance so that the window statistic is approximately -distributed under the null, yielding a closed-form, label-free threshold , and (iii) fuses several window scales by robust per-scale standardization and a max rule that encodes “anomalous at some scale”, with classical multiple-testing combinations (Sidak/Fisher) analyzed as calibrated alternatives. The entire pipeline consumes exactly one prior: an upper bound on the false alarm rate. On three real NAB streams and two controlled synthetic benchmarks, BandSAD attains the best mean point-adjusted among six established baselines (Isolation Forest, PCA, Spectral Residual, AE, LSTM-AE, USAD) and the best AUC-PR on the strongly periodic stream, while remaining competitive on the others; more importantly, it is the only method whose decision threshold needs no labeled anomalies—the calibrated recovers up to of the oracle-threshold on streams with adequate anomaly prevalence, and Q–Q analysis confirms the null in the body of the distribution. Ablations show that each component—band decoupling, out-of-sample MAD whitening, and multi-scale fusion—contributes measurably to the final performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Univariate time series, anomaly detection, frequency band decoupling, Fourier transform, autoencoder, label-free threshold, false alarm rate, chi-square distribution, multi-scale fusion, time series forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_zHLqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes BandSAD, a forecasting-based method for univariate time-series anomaly detection without labeled anomalies. A learnable Fourier-domain partition-of-unity mask decomposes the forecasting target into multiple components, which are predicted by a shared convolutional–MLP network. Component-wise forecasting residuals are normalized using held-out MAD-based scale estimates and combined into a window statistic. Under an idealized independent Gaussian residual model, the statistic admits a chi-square threshold. The deployed algorithm instead uses an empirical calibration quantile after robust standardization and maximum fusion across three temporal scales.

Critical strengths

The learnable Fourier-domain target decomposition is an interesting methodological design. Its differentiable partition-of-unity masks allow the decomposition and forecaster to be trained jointly while preserving exact reconstruction of the original target, offering a flexible framework for component-wise forecasting and anomaly scoring.

Critical weaknesses
  1. The claimed statistical guarantee is not established.
    Proposition 1 assumes independent Gaussian residuals, but the soft Fourier masks guarantee only a partition of unity, not orthogonality or residual independence. The paper does not sufficiently justify these assumptions for the actual residuals, leaving its central claim of exact false-alarm control unsupported.

  2. The empirical evidence does not establish a convincing advantage.
    Comparisons are limited to six established baselines, without recent competitive methods. Average AUC-PR and point-wise F1 are below those of the standard AE baseline. Although average PA-F1 is highest, inconsistencies in the reported PA-F1 and false-positive rates undermine confidence in the evaluation. Moreover, learned masks do not outperform fixed masks in the reported ablations.

Questions for authors
  1. What justifies residual independence and exact false-alarm control, given that the soft Fourier masks do not enforce orthogonality?

  2. How is PA-F1 computed, and why is it substantially lower than point-wise F1 at the same label-free threshold in Table 2?

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

The paper presents a subband forecasting model for label-free time-series anomaly detection. It normalizes prediction errors using robust variance estimates and claims a Chi-squared-based false-alarm guarantee.

Critical strengths

The problem the authors are addressing of setting anomaly detection threshould without labeled anomalies is very relevant. The presented approach of subband forecasting, robust normalization, and multi-scale detection is sensible and sounds promising.

Critical weaknesses

Anomaly detection requires an operational characterization of the nominal data-generating process, whether through a density model, decision boundary, predictive distribution, or an appropriately calibrated anomaly statistic. BandSAD learns a conditional-mean forecaster and estimates robust per-band residual scales, but neither operation establishes the distribution of normal forecasting errors. This becomes very apparent when normal behavior is multimodal, heavy-tailed, or impulsive, etc. e.g., in a compound-Poisson process large and unpredictable residuals can be ordinary realizations of the nominal distribution. MAD-based normalization does not identify such behavior as normal, it may instead amplify the corresponding anomaly scores. More generally, the fact that a residual is unpredictable does not imply that it is anomalous, and a predictable trajectory need not be normal.

The paper's own findings undermine the practical applicability of its central theoretical claim. In Section 3.3 (Remark 1, p. 4), the authors report residual kurtosis of approximately 15 (versus 3 under Gaussianity) and acknowledge that the theoretical Chi-squared threshold yields false-alarm rates approximately twice the nominal level, with substantially larger deviations under distribution drift (Section 4.4, Table 3). Consequently, the actual detector replaces the analytical threshold with an empirical calibration quantile (Section 3.3, Remark 1; Section 3.4; Algorithm 1, line 14).

Questions for authors

What property of the proposed learning objective ensures that the distribution of standardized forecasting errors represents the full range of nominal behavior, including multimodal, heavy-tailed, and impulsive processes? If no such property exists, how to motivate the Chi-squared null distribution, and what is the advantage of the proposed calibration over applying a standard empirical quantile to any other forecasting-based anomaly score?

Reviewer_ZB2jOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes BandSAD, which uses a learnable partition-of-unity in the Fourier domain to decompose the future segment into K frequency bands. It forecasts each band separately and applies MAD-based whitening to the residuals from a held-out calibration segment. Under an idealized independent Gaussian residual assumption, the authors derive a chi-square window statistic and a threshold parameterized by the desired significance level. The detector is then run at three window scales, robustly standardized per scale, and fused by a max rule. Experiments compare BandSAD with six baselines on three NAB streams and two synthetic streams, reporting oracle-threshold results, label-free results, Q–Q analysis, and component ablations.

Critical strengths
  1. A clear deployment objective. The paper makes threshold selection without anomaly labels its central problem and reports oracle and label-free modes separately.
  2. Interpretable components and failure modes. The band, MAD-whitening, multi-scale, and theoretical-versus-empirical threshold components are ablated or compared, and the authors acknowledge drift, overlapping windows, and heavy tails.
Critical weaknesses
  1. The independence argument in the chi-square derivation is a central theoretical error. The soft masks in Eq. 1 satisfy only ∑kmk=1\sum_k m_k=1; they do not imply that mkUm_kU and mk′Um_{k'}U are orthogonal. Parseval's theorem and partition-of-unity therefore do not establish independence across bands. Horizon steps and stride-1 windows are also dependent, and σ is estimated from a finite calibration sample. Consequently, the claimed “exact nominal false-alarm rate” does not apply to the actual detector. The “FPR” in Table 3 is actually the fraction of all test points, including anomalous points, above threshold; it is only a proxy for false alarms and cannot replace a clean-only FPR evaluation. This highlights that the nominal false-positive rate is not directly validated.
  2. The label-free evidence and comparison protocol do not support the deployment claims. In Table 2, label-free PA-F1 is very low on TravelTime, Syn-Mix, Syn-Chp, and parts of Ambient Temperature. The main comparison uses oracle thresholds for every method, BandSAD uses three seeds while baselines use seed 0, and the evaluation contains only three NAB streams and two synthetic streams. There is no independent clean-test FPR, multi-seed confidence interval, drift-segment analysis, or conformal/block-bootstrap baseline. The complete algorithm also replaces the closed-form threshold with an empirical calibration quantile, so the practical role of the theoretical threshold needs to be reframed.
Questions for authors
  1. Please report residual covariances and independence diagnostics across bands, horizon steps, and adjacent windows. If independence does not hold, please reformulate the contribution as an empirically calibrated detector and report FPR confidence intervals on fully unseen clean-drift streams.
  2. Please use the same label-free threshold protocol for all methods, report multi-seed means, standard deviations, and independent clean-test FPR, and clarify whether the training/calibration segments are truly anomaly-free and temporally leakage-free.
  3. Please compare block-bootstrap, conformal calibration, and probabilistic forecasting baselines, and report PR/FPR curves under different anomaly prevalences, durations, and distribution shifts. The current synthetic label-free results are very low and require an explanation of their deployment value.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.