acceptodds
Under review as a conference paper at ICLR 2027

FLARE: Factor-Loading-Break-Triggered Adaptive Readout for High-Dimensional Time Series under Regime Shift

Abstract

Time-series forecasters can degrade sharply when the shared low-rank structure of a high-dimensional panel changes across regimes. Yet common remedies such as per-series instance normalization primarily target marginal distribution shifts and can therefore miss changes in the joint dependence structure. To address this challenge, we formulate regime shifts as low-rank structural breaks, covering both span-changing shifts in the loading space and span-preserving changes in factor strength. Building on this formulation, we introduce FLARE, an online detector that uses a whitened amortized representation and a single operator statistic to jointly capture subspace geometry and factor strength. Each detected break then triggers lightweight readout-only adaptation while the forecasting backbone remains frozen. Theoretically, we establish a weak-factor detectability boundary and a detection-delay guarantee, and show that whitening lowers the detection threshold relative to window-wise PCA. Empirically, FLARE detects weak-factor breaks missed by classical factor detectors, localizes a real market regime shift, and matches the forecasting performance of always-on test-time adaptation at a fraction of the update cost.

Primary area:
learning on time series and dynamical systems
Keywords:
change point, time series, online learning, factor model

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_9CW4OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FLARE, an online detect-then-adapt framework for high dimensional time series under low-rank structural breaks. A whitened amortized loading encoder estimates window-specific eigenbases and eigenvalues. A unified statistic compares rank-KK truncated whitened-covariance operators to capture both span-changing and identifiable span-preserving breaks. An ARL-calibrated, self-normalized CUSUM triggers readout-only adaptation while keeping the forecasting backbone frozen. The paper presents weak factor detectability and detection-delay guarantees, synthetic detection comparisons and ablations, regime localization results on two economic panels, and a synthetic forecasting comparison reporting similar post-break MSE to continuous adaptation with substantially fewer updates.

Critical strengths

1. The unified operator statistic addresses an important limitation of geometry-only or spectrum-only monitoring. Comparing rank-truncated whitened-covariance operators gives a coherent representation of both span-changing and identifiable span-preserving breaks. Equation (4) has a valid closed-form expansion, and the rank-one identity T=(wL−wR)2+2wLwRsin⁡2θT=(w_L-w_R)^2+2w_Lw_R\sin^2\theta makes the complementary contributions of spectral strength and subspace geometry explicit. This is a meaningful methodological feature because either component alone can miss structural changes relevant to forecasting.

2. ARL matched synthetic comparisons provide evidence of weak-factor sensitivity. Calibrating the competing detectors to a common target ARL and reporting realized null ARLs is an appropriate comparison strategy. In the reported benchmark, FLARE reaches both recall and F1 boundaries at (δ=0.50\delta=0.50), whereas the strongest external baseline reaches a recall boundary of (0.70). The statistic ablations also demonstrate complementary sensitivities to span-changing and span-preserving breaks. These results support further investigation of the proposed representation.

Critical weaknesses

1. The stated weak-factor detection boundary is inconsistent.

Theorem 5(i) states that consistent detection is impossible when θ2=o(ρp,h2).\theta^2=o(\rho_{p,h}^2). However, the achievability conditions in Theorem 5(ii) can simultaneously hold in this regime. For example, take θ2=ρp,h2h−1/2,log⁡γ=h1/4,Δ=h2.\theta^2=\rho_{p,h}^2h^{-1/2}, \qquad\log\gamma=h^{1/4},\qquad\Delta=h^2.

Then (θ2=o(ρp,h2))(\theta^2=o(\rho_{p,h}^2)), while ρp,h2θ−2log⁡γ=h3/4=o(h)\rho_{p,h}^2\theta^{-2}\log\gamma=h^{3/4}=o(h) and θ2ρp,h2log⁡γ/Δ=h5/4→∞.\frac{\theta^2}{\rho_{p,h}^2\log\gamma/\Delta}=h^{5/4}\rightarrow\infty.

Thus, the stated impossibility and achievability results apply to the same parameter sequence. Appendix B.2 explicitly uses the separation condition (θ≫ρp,h\theta\gg\rho_{p,h}), but this condition is absent from the corresponding theorem statements. Adding it may remove the contradiction, but would require revising and reassessing the claimed detection boundary. This directly affects a central theoretical contribution.

2. The assumptions and analysis do not establish the guarantees for the deployed unified statistic.

The weak-factor guarantees depend on the amortized subspace rate ρp,h=p−δ/2h−1/2\rho_{p,h}=p^{-\delta/2}h^{-1/2}, but the manuscript does not provide sufficient conditions on the encoder, its training information, or its structural priors to establish this rate. The covariance perturbation argument in Appendix B.1 does not adequately justify the asserted rate or the whitening-versus-PCA ordering. Moreover, the proof primarily analyzes the angle statistic (DtD_t), whereas the deployed (Tlog⁡T_{\log}) also contains spectral estimation error. Even with a perfectly known loading direction, window-specific log-eigenvalues in a Gaussian one-factor model can fluctuate at order (h−1/2h^{-1/2}), producing an order-(h−1h^{-1}) contribution to the operator distance under the null. This variation is not controlled by the angle noise scale (p−δh−1p^{-\delta}h^{-1}). HAC self-normalization may accommodate the actual variation, but the normalized drift and detection delay would then require additional analysis. Lemma B.3 does not supply that analysis. Consequently, correcting the overlapping boundary conditions alone would not establish the advertised guarantees for FLARE.

3. The detection experiments do not isolate the claimed whitening advantage or establish the headline performance under the proposed unknown-rank procedure.

The comparison between per-window PCA and the whitened amortized encoder changes whitening, amortization, and the estimator class together. It therefore supports the combined representation, but does not establish how much of the improvement is attributable to whitening. This distinction matters because lowering the weak-factor detectability threshold through whitening is a central claim. The working-rank results raise an additional concern about the headline operating point. In Table C.2, the reported recall/F1 boundaries of (0.50/0.50) match the fixed-(K=4K=4) setting, whereas the proposed plug-in rule (K=r^+1K=\widehat r+1) yields (0.60/0.70). The manuscript does not sufficiently reconcile these results with the principal comparisons. The empirical advantage should be established under the stated unknown-number-of-factors setting, with the contribution of whitening distinguished from that of the learned encoder.

4. The forecasting evaluation does not establish the benefit of break-triggered readout-only adaptation or an end-to-end accuracy–adaptation-cost advantage.

The main forecasting evidence consists of synthetic post-break MSE values from an unspecified lightweight temporal backbone: 2.490 for the frozen model, 2.356 for FLARE, and 2.352 for TAFAS. This comparison can change both the adaptation module and the update schedule, so it does not isolate the benefit of using structural-break detections as adaptation triggers. The real-panel results primarily report localized dates and sparse updates, rather than comparative post-break forecasting accuracy and online detection quality. They therefore provide limited support for the practical forecasting contribution or its backbone-agnostic applicability. Finally, one update versus 24–30 updates establishes a reduction in update frequency, but does not account for the continuous cost of the amortized encoder, whitening, and detector. Without a controlled comparison using the same backbone and readout, together with end-to-end runtime measurements, the claimed accuracy–adaptation-cost trade-off remains insufficiently substantiated.

Questions for authors
  1. How should Theorem 5 be revised to eliminate the overlap demonstrated above? After including the necessary separation conditions, what detection boundary and delay guarantee remain valid?
  2. What training information and assumptions establish the amortized subspace rate in Assumption 2? What result controls spectral estimation error and transfers the angle-based analysis to Tlog⁡T_{\log}?
  3. Which rank-selection procedure is used in Tables 1 and C.1? Does the reported advantage persist when the plug-in (K=r^+1K=\widehat r+1) rule is used consistently? Can the contribution of whitening be distinguished from that of amortization and the encoder class?
Reviewer_6aK3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FLARE, an online method for detecting structural changes in high-dimensional time series and adapting forecasting models accordingly. It uses a whitened encoder and a unified statistic to detect changes in both factor loading spaces and factor strengths. When a change is detected, FLARE updates only the prediction head while keeping the backbone frozen. The paper also provides theoretical results on weak-factor detectability and detection delay. Experiments on synthetic and real-world datasets evaluate the detection performance and adaptation efficiency.

Critical strengths
  1. The paper addresses a meaningful problem in high-dimensional time series forecasting, focusing on structural changes in factor dependencies rather than marginal distribution shifts.

  2. The proposed unified statistic provides a reasonable way to detect different types of structural breaks. The experiments show promising detection performance and reduced adaptation costs compared with existing methods.

Critical weaknesses

1.The convergence rate of the learned encoder is assumed rather than proven. In addition, the proofs of the detection boundary and false-alarm guarantees contain gaps that need further clarification.

  1. The method combines whitening, covariance-based change detection, and event-triggered adaptation, which are largely established techniques. Although the unified statistic is reasonable, its novelty and advantages over existing approaches are not sufficiently demonstrated.
Questions for authors

see above

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

This paper proposes FLARE as a framework for high-dimensional time-series forecasting under regime shifts characterized by low-rank structure change, which adapts lightweight readout of a frozen backbone. Specifically, FLARE follows a regime-shift detect and adapt structure, and the authors present the following elements as the main components of FLARE.

(1) Representation Whitening – introduced to mitigate the effect of the heteroskedastic nature of the ete_t component on the representation.

(2) Amortized loading encoder – circumvents limitations of the existing PCA encoder (computational efficiency, non-differentiability, etc.) and supports joint optimization with the detector.

(3) A single-calibrated discriminator – a discriminator that can detect both span-preserving and span-changing breaks.

The authors claim that FLARE guarantees weak-factor detection, and the paper's empirical results suggest that FLARE outperforms existing methods in weak-factor detection.

Critical strengths
  1. Clear identification of research gap.
  • This paper targets a specific gap that existing methods have overlooked: detecting regime shifts driven by weak factors in settings where the shared low-rank structure changes across regimes.
  1. Formal bridging between time-series forecasting test-time adaptation and regime shifts.
  • FLARE characterizes regime shifts arising in time-series forecasting tasks through a low-rank structure and proposes a formal framework including weak-factor detector.
Critical weaknesses
  1. Lack of verification and mechanism-level analysis.
  • Each component introduced by the authors has a plausible motivation, but for some components the mechanism has not been sufficiently verified.

(1) Representation Whitening: Partially supported by Subspace-CUSUM raw vs. whitened in Table 1, which is a mixed effect of encoder selection. No measurement of principal angles such as the change in sin⁡Θ\sin \Theta before and after whitening provided.

(2) Amortized loading encoder: The reported ablation (Sec. 5.2 and App. C.2) changes whitening and the encoder at the same time, and the ablation isolating encoder selection (e.g., amortized loading encoder vs. PCA encoder) is not presented. Furthermore, a more specific comparison on output basis is required (please see Weakness 2).

(3) A single-calibrated discriminator: Sufficiently supported by Table 2, C.3, and C.4. However, several design choices, such as the log weighting of eigenvalues are made without enough sensitivity analysis.

  1. Some concerns on amortized loading encoder
  • The amortized loading encoder is learned in an autoencoder manner, and it is well known that a linear autoencoder trained with a reconstruction loss recovers the principal subspace. Since the loss depends on the encoder only through its output basis U^t\hat{U}_t, for which the optimum is per-window PCA, the optimal encoder is expected to reproduce per-window PCA regardless of the encoder architecture. Given that the authors proposed the amortized loading encoder as a substitute for the PCA encoder, a more specific analysis of the difference between the bases produced by the proposed encoder and by the PCA encoder would make the proposed method more convincing.
  • There are no details on the encoder, e.g., the encoder architecture, the masking rule, or the training data.
  1. Insufficient experiments on real-world datasets.
  • Most of the experiments and analyses are based on a synthetic dataset generated by the authors and designed to reflect weak-factor detection. As FLARE mainly focuses on weak-factor detection, this dataset may inherently work in FLARE's favor. To make FLARE's contribution clearer, more experiments on real datasets should also be conducted.
Questions for authors
  1. Please provide mechanism-level evidence (not just final metric comparisons) for (1) representation whitening and (2) the amortized loading encoder. I wonder whether these proposed components really deal with the problem that the motivation points out.

  2. Please provide a basis comparison (e.g., measurement of grassmann distance, etc.) between the traditional PCA encoder and the amortized loading encoder. I expect the basis produced by the amortized loading encoder to be close to (nearly coincide with) that of the PCA encoder, but I would like to confirm through actual measurement whether this is really so.

  3. Please provide empirical results and an ablation study on real-world datasets, including comparison with baseline methods. Although the paper provides several results and ablations, these are mainly based on an author-generated synthetic dataset.

Reviewer_KJG7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper claims to propose a method for detecting structural breaks in multidimensional time series data; this method not only detects structural breaks but also adapts the data for subsequent forecasting. Furthermore, the paper claims that “However, these methods primarily target statistical detection or break localization rather than using detected breaks to trigger forecasting adaptation, and many rely on sufficiently strong factors for reliable eigenspace estimation. These limitations leave factor change-point methods poorly suited to online forecasting, particularly in weak-factor regimes, where structural changes must be detected promptly and translated into adaptation decisions. ”—in other words, no previous study has addressed this issue. However, the experiments in this paper evaluated predictive performance only on synthetic data and used mean squared error (MSE) as the sole evaluation metric. It is also worth noting that the authors appear to lack an understanding of online machine learning. This paradigm—in which a concept drift detector identifies concept drift and then adapts and updates the model—has long been established.

Critical strengths

The author proposes a new method and provides some theoretical analysis.

Critical weaknesses

Firstly, the author’s review of the relevant literature is insufficient. Whilst in finance this might be referred to as a ‘structural break’—and in the context of multitemporal time series data as ‘panel data’—within the computer science community, such a structural break can be viewed as an anomaly detection problem involving contextual anomalies, or as a concept drift detection problem within online machine learning. Secondly, the so-called existing work focuses solely on how to detect such breaks, without addressing how to adapt the model once detection has taken place, In fact, in online machine learning, a paradigm emerged long ago whereby concept drift is first detected using a concept drift detector, followed by online model updates. This is not a ‘research gap’ at all, but rather a topic that has been thoroughly discussed. Examples include Hoeffding trees with ADWIN, amongst others, Please read Chen et al., Contextual Time Series Change Detection,SDM 2013. Zhao & Shen, Proactive Model Adaptation Against Concept Drift for Online Time Series Forecasting, KDD 2025. Zhang et al., OneNet: Enhancing Time Series Forecasting Models under Concept Drift by Online Ensembling, NeurIPS 2023 Second, the author’s experimental design does not align with the claims made in his paper. If the motivation of his paper is to demonstrate that, in addition to detecting structural breaks, he can also enhance the adaptability of predictive models—and if the focus should be on the adaptability of predictive models—then he should have compared his approach with other methods for adapting predictive models. For example, in online machine learning, once a concept drift detector detects concept drift, it triggers an update in the online machine learning method; This paper should have compared a wider range of concept drift detectors, as well as context anomaly detectors, to demonstrate how, after detecting concept drift, the model can achieve better performance. However, the data used here for prediction performance consists solely of synthetic data, and the scale of the dataset is not clearly disclosed. Furthermore, in traditional econometrics, researchers would typically conduct significance tests; numerically speaking, a difference of just one digit after the decimal point—between 2.352 and 2.356—makes it difficult to conclude that there is any meaningful difference. Additionally, the authors have not disclosed whether the results represent the average of multiple experiments with multiple seeds, nor have they provided standard deviations, error upper bounds, or lower bounds, the reliability of this experiment has not been substantiated. The third point concerns the choice of baseline in the experimental design. The most recent method currently available is a 2018 preprint v1 Subspace-CUSUM, whilst many of the latest context-aware anomaly detection methods are overlooked, as are methods for concept drift detection; furthermore, no direct reference is provided for the source of the other baselines.

The fourth point concerns the writing of this paper: many aspects are not explained clearly, such as the specific details of the model and the choice of backbone for the model; moreover, the logic of the writing is very difficult to follow. It is written in a style that resembles neither a traditional econometrics paper nor a paper for a computer science conference.

Questions for authors

Since the article has already mentioned online machine learning, why not consider the framework and methodology of online machine learning-based concept drift detectors for comparison? Why use only MSE, rather than other metrics such as MAE? And why not provide a significance analysis or the error ranges from multiple experimental runs, amongst other things?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.