acceptodds
Under review as a conference paper at ICLR 2027

Decay-Weighted Survivability Estimators: A Nonparametric Class of Regime Detectors with Bounded Forgetting and Reactivation

Abstract

Tracking regime shifts in non-stationary time series is difficult because standard filtering and changepoint detection algorithms typically discard inactive hypotheses to remain computationally efficient. When the environment shifts abruptly, or reverts to a regime seen before, these models are forced to reconstruct state estimates from limited recent observations. We introduce Decay-Weighted Survivability Estimators (DWSE), a class of nonparametric regime detectors that instead retain inactive hypotheses through a decay-plus-floor trust recurrence: rather than deleting a hypothesis, its trust decays toward a small but nonzero floor, allowing rapid reactivation if the regime it describes recurs. This guarantee is distinct in kind from forgetting-factor results in adaptive filtering, which bound convergence of a parameter estimate; DWSE instead bounds the dynamics of hypothesis selection itself. We instantiate the class in three ways: DWSE-Mean and DWSE-Vol, targeted to shifts in the first and second moments respectively, and DWSE-Rank, an omnibus detector sensitive to distributional change more broadly.We prove finite-sample bounds on hypothesis forgetting time (), reactivation time (), and worst-case detection delay (), alongside a non-asymptotic concentration guarantee. Under stable regime conditions, we further show the centered trust process satisfies a functional central limit theorem, converging weakly to a scaled Brownian motion, giving a principled basis for threshold calibration. On Bitcoin and non-stationary climate data, all three DWSE instances achieve strong recall on labelled regime shifts, performing comparably to or better than classical baselines such as CUSUM, EWMA control charts, directional change, Kalman filtering, and particle filtering, all with the same fixed parameters across domains.

Primary area:
learning on time series and dynamical systems
Keywords:
Probabilistic Methods, Time Series and Sequence Models, Theory, Applications to physical sciences

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_P2B6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Decay-Weighted Survivability Estimators (DWSE), a nonparametric framework for regime-change detection. It maintains a fixed set of hypotheses and updates their trust using exponentially decayed compatibility scores, without specifying a parametric observation distribution. A positive trust floor is intended to preserve inactive hypotheses so they can be reused when previous regimes recur. The paper provides theoretical forgetting and recovery bounds and evaluates the method on financial and climate datasets.

Critical strengths
  1. DWSE does not require specifying Gaussian, Student-(t), or other parametric observation likelihoods. Maintaining reusable hypotheses is also an appealing design for recurring regimes.

  2. The authors report weaker performance on some datasets and include ablations that do not consistently support the proposed trust-floor mechanism, rather than presenting only favorable outcomes.

Critical weaknesses
  1. The trust floor contribution is unclear: without the floor, the DWSE trust update is mathematically equivalent, after rescaling, to an EWMA of hypothesis compatibility scores. Moreover, retained hypotheses can recover even when (epsilon=0). The synthetic deletion experiment compares against permanently removed hypotheses rather than equally retained zero-floor hypotheses, while the real-data floor ablation shows little difference. This leaves the floor's claimed contribution insufficiently established.

  2. Insufficient evidence for robustness and computational efficiency. Appendix D shows substantial sensitivity to the detection threshold, but its calibration is not adequately justified. The paper also motivates its approach partly through avoiding hypothesis deletion, yet provides no runtime or computational-cost comparison with competing methods.

Questions for authors
  1. How does DWSE differ from EWMA-based expert tracking or fixed-share methods? Could the authors compare against a multi-hypothesis EWMA?

  2. How were the default values (Z=2) and (gamma=0.8) selected? Appendix D suggests substantial threshold sensitivity. Could the authors provide a finer threshold sweep, a recall–false-alarm trade-off curve, and practical guidance for tuning, e.g. gamma?

  3. Is the implementation online or can only be used in offline setup? Is all initialization and adaptive window selection performed causally, without access to future observations? What are the runtime and memory costs compared with methods that prune inactive hypotheses?

  4. Can the authors provide intuition for the datasets or regimes where DWSE performs poorly? Are there identifiable conditions under which users should avoid using it?

  5. Given the paper's emphasis on nonparametric regime detection, could the authors compare DWSE with an online or sliding-window MMD-based change-point detector under the same evaluation protocol? This would help establish its advantages over existing distribution-free alternatives.

  6. The claim of being distribution-free should be qualified. While the update does not require a parametric likelihood, to what extent do the concentration and asymptotic results depend on assumptions such as independence and stationarity?

  7. The paper would also benefit from a methodological diagram, clearer definitions of parameters when first introduced (in introduction), and plots showing detection signals and alarms alongside the observed time series. These additions would make the empirical results easier to interpret.

Reviewer_FwTeOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes tracking a fixed collection of candidates through exponentially down weighted compatibility scores for online change point detection. I have substantial concerns about the exposition, the statistical interpretation of the candidates, and the significance of the theoretical contribution.

Critical strengths

None

Critical weaknesses
  1. The introduction does not clearly formulate the changepoint problem. It should explain what feature of the data-generating process changes—such as the mean, variance, or distribution—what assumptions describe observations within a regime, and whether the objective is to declare an online alarm or identify the current regime. Instead, terms such as “environment shifts,” “hypothesis,” “trust,” and “survivability” appear before their statistical meanings are established. The first mathematical display also uses notation before its meaning is clearly explained. A concrete change-point example and definitions at first use would make the method and its objectives much easier to understand. The current version is very difficulty to read even for researchers who is familiar with the change point analysis.

  2. The candidate hypotheses are not adequately connected to segment distributions or parameters. Each candidate appears to define a rule that scores compatibility with an observation. For example, a DWSE-Mean candidate receives support when returns fall within a band around its proposed level. Maximizing expected support therefore identifies a band containing high probability mass, which need not identify the segment mean. The concentration result establishes selection of the candidate with greatest expected compatibility, under an assumed separation condition. An additional identification argument is needed to interpret that candidate as the correct regime. The authors should specify how candidates represent statistical regimes and explain which changes in the underlying process produce detectable changes in their scores.

  3. The methodological and theoretical advances remain unclear. The update combines exponential discounting with a positive score floor, while retaining candidates for later recovery has precedents in expert tracking. The unchanged no-floor ablation also leaves the floor’s practical contribution unclear. The forgetting and recovery bounds describe elementary filter responses to prescribed compatibility sequences; concentration uses standard inequalities; and the functional CLT applies standard linear-process theory under floor inactivity. From a change-point perspective, these results do not establish a false-alarm versus detection-delay guarantee for the implemented detector under a specified pre-change and post-change model. The authors should clarify what capability the method adds and what new statistical guarantee supports that capability.

Questions for authors

None

Reviewer_JrU2OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper propose a method to detect regimes in non-stationary time series by maintaining trust score under a recurrence without decaying until zero. To update the trust scoreThe paper uses compatibility score which computes from different statistics including mean, volatility, and rank. The paper evaluates the method with three different datasets.

Critical strengths

I find the current problem setting can be interesting. However, in the current form, the weaknesses outweight the strengths.

Critical weaknesses
  1. The paper presentation may needs substantial revisions. First, the introduction provide many math notations without having clear setup like Tt(h)T_t(h), TforgetT_{forget} and TcoverT_{cover}. Second, I find the writing is natural and hard to parse. Third, the paper is not in a correct format for reviews with "Published as a conference paper at ICLR 2027".

  2. The reference contains limited number of citations which is likely insufficient literature reviews.

  3. Tuning parameter γ\gamma be problematic. Clearly, Table 6 shows that the parameter is sensitive. Slightly different values of γ\gamma can result in a big difference in the performance.

Questions for authors
  1. What is false-alarm rate (FAR) used in the experiments? How can we justify it without a regime ground truth in real-world data?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.