{"id":"77a77ebf-73f4-4a11-a3a7-9e63c3f99f41","arxiv_id":"2509.00585","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MOPED uses moving sums of tail pairwise dependence matrix estimates to detect multiple change points in extremal dependence, with a multiscale variant that pools thresholds and bandwidths.","lead":"This paper introduces MOPED, a nonparametric algorithm that detects abrupt changes in the tail dependence of multivariate time series. It is tested in simulations and applied to neonatal EEG recordings, where detected change points align with clinician-annotated seizures.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MOPED's EEG seizure findings may be driven by nonstationary marginal scale changes rather than changes in tail dependence.","rationale":"The reader's weakest assumption combines serial dependence and marginal stationarity, but the single most load-bearing condition for the EEG claim is marginal stationarity. Seizures are known to involve amplitude changes, so the empirical rank transform cannot be assumed to remove marginal effects. The proposed test isolates this mechanism by holding the copula fixed and only changing marginal scale; a positive result would directly falsify the interpretation of the EEG change points as tail-dependence changes. A negative result would leave the paper's central application intact on this axis, leaving serial dependence as the main remaining caveat. Because the paper's own assumptions are explicit, this is a missing verification rather than an internal inconsistency, and the conditional verdict is appropriate.","tokens_in":23935,"tokens_out":8457,"duration_ms":109160,"concrete_test":"Generate n=5000 bivariate observations from a t-copula with ν=3 and fixed correlation ρ=0.5 (constant tail dependence throughout), with margins that change scale at t=2500: e.g., multiply both components by a factor of 3 for t>2500. Apply the same preprocessing as Section 4.1: global empirical rank transform to Pareto(2) margins, then run MOPED with G=1500, k=0.1G and permutation α=0.05. If MOPED returns a significant change point near 2500, the method cannot separate marginal nonstationarity from dependence change, and the EEG application is confounded. If no change point is found, this particular concern is mitigated (though serial dependence remains).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central methodological claim is conditional on stationary marginal distributions (Section 1: 'hereafter assume that the marginal distributions are stationary'), but the EEG application never checks this. Section 4.1 standardizes the full series to Pareto(2) margins via a single empirical rank transform. If seizure activity changes marginal scale or tail heaviness, this global transform does not yield identically distributed Pareto margins over time; observations from high-amplitude seizure periods become systematically larger after ranking. Because the TPDM estimator (Eq. 3) is driven by local radial exceedances, such marginal changes can alter the estimated TPDM even when the copula is constant. MOPED would then declare change points at seizure boundaries that are changes in marginal behavior, not in pairwise extremal dependence. This undermines the abstract's claim that MOPED 'identifies significant structural changes in the extremal dependence' of EEG signals during seizures, and more generally the interpretability of MOPED when the stationarity assumption is violated. The serial-dependence issue raised by the reader compounds this but is not the primary confound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MOPED, a nonparametric moving-sum (MOSUM) procedure for detecting multiple change points in the tail pairwise dependence matrix (TPDM) of a multivariate regularly varying time series. The detector compares TPDM estimates in left and right windows via a Frobenius-norm statistic, with significance thresholds obtained by permutation. A multiscale, multi-threshold extension pools change point estimates across bandwidths G and extremal threshold orders k. The method is evaluated in simulations against E-divisive and the parametric method of Hazra and Bose (2025), and applied to neonatal EEG recordings, where the authors report change points near annotated seizure activity.","tokens_in":24186,"tokens_out":3193,"duration_ms":43313,"significance":"If the central claims hold, MOPED fills a real gap: it provides a computationally feasible, nonparametric, multiple-change-point procedure for tail dependence that scales to dimensions where existing extremal tests are unavailable or computationally prohibitive. The simulation design in Scenario 2 is a particular strength: it demonstrates that MOPED can detect changes in extremal dependence class (asymptotic dependence vs. independence) even when the correlation structure is unchanged, a case where general-purpose methods like E-divisive struggle. The paper is also accompanied by an R package, and the simulation results are reported in reproducible detail. However, the strongest claims—especially the EEG interpretation—depend on assumptions of stationary marginal distributions and serial independence that are not verified, and the paper itself acknowledges that asymptotic theory is absent. These issues limit the current support for the abstract's wording that MOPED 'identifies significant structural changes in the extremal dependence' during seizures.","major_comments":[{"comment":"The EEG analysis standardises the full series to Pareto(2) margins using a single empirical rank transform (Section 4.1), while the method assumes stationary margins (Section 1). If seizure activity changes the marginal scale or tail heaviness, the global transform does not produce identically distributed Pareto margins over time. Because the TPDM estimator in Eq. (3) is driven by radial exceedances, such marginal changes can alter the estimated TPDM even when the copula is constant. MOPED would then declare change points at seizure boundaries that reflect changes in marginal behaviour, not pairwise extremal dependence. The paper should either test marginal stationarity in the EEG data (e.g., via a local or sliding-window marginal fit), apply segment-wise marginal transformations, or explicitly reinterpret the empirical findings as 'changes in the tail region' rather than 'changes in ext","section":"Section 4.1 and Eq. (3)"},{"comment":"The permutation threshold in Section 2.5 is justified only 'under the assumption of independence of the observations.' In the EEG application, serial dependence is addressed by subsampling every 256th observation, but this does not remove potential extremal dependence at the sampled scale, and residual serial dependence can inflate the permutation null and produce spurious change points. The paper should discuss this limitation and preferably use a block permutation or dependent bootstrap scheme that preserves local dependence, or provide evidence that the subsampled EEG series passes a test for serial independence at extreme levels.","section":"Section 2.5 and Section 4.1"},{"comment":"The paper explicitly states that 'the theoretical treatment of the MOPED algorithm, i.e., the asymptotic characterisation of the MOSUM test statistic and convergence rates for the change point estimators, remains an avenue for further work.' Without such theory, there is no formal justification that the detector in Eq. (6) is consistent, that the η-criterion in Eq. (8) recovers the true number and locations of change points, or that the permutation p-values in Eq. (9) are calibrated under the model in Eq. (4). The empirical evidence is supportive, but the abstract and discussion should temper the claim that MOPED 'identifies' change points, and state clearly that the method is currently justified by simulation rather than asymptotic guarantee.","section":"Section 5"},{"comment":"The multiscale, multi-threshold variant (MMMOPED) is acknowledged in Section 3.1 to be 'less conservative' and prone to spurious estimates. The simulations quantify this: for q=0, MMMOPED returns the correct number of change points in only 56.8% of replications (Table 1, d=2, ρ=0.2), versus 89.7% for fixed MOPED, and it frequently reports false positives (13.3% with bq-q ≥ 2). Given that the paper recommends MMMOPED for 'applications where the accuracy of change point estimates is more important than testing,' the manuscript should provide clearer guidance on when the false-positive rate is acceptable, and should not present MMMOPED as a universally superior alternative without qualification.","section":"Table 1 and Section 3.1"}],"minor_comments":[{"comment":"Typo: 'timesseries' should be 'time series'.","section":"Section 2.5"},{"comment":"Typo: 'eprformance' should be 'performance'.","section":"Section 2.6"},{"comment":"In the pseudocode, 'Add bk to bC' should read 'Add bτ to bC'.","section":"Algorithm 3"},{"comment":"The sentence 'Observations of {Xi,t} are plotted against time t...' is repeated verbatim; one occurrence should be deleted.","section":"Figure 6 caption"},{"comment":"The phrase 'local estimates of the TDPM' should be 'TPDM' for consistency with the abbreviation used elsewhere.","section":"Section 2.4"},{"comment":"The abstract states that MOPED 'identifies significant structural changes in the extremal dependence of the signals when the subjects undergo seizures.' Given the marginal-stationarity assumption and the absence of a formal test for marginal stationarity in the EEG data, the wording should be softened (e.g., 'is consistent with changes in extremal dependence') unless additional validation is provided.","section":"Abstract and Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid contribution to an active area, and the simulation study is convincing enough to justify publication after revision. The main risk is that the EEG application overstates the method's ability to isolate changes in tail dependence when marginal nonstationarity is a plausible alternative explanation. The absence of asymptotic theory is a known limitation and is honestly acknowledged, but it should be more prominent in the paper's framing. I do not see evidence of circularity or problematic citation practices; the use of the authors' own prior work is appropriate and openly referenced."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core method is real: applying MOSUM to TPDM estimates is a sensible, new combination, and the multiscale/multi-threshold merging addresses a genuine tuning problem. Second, the EEG application is not clean evidence for 'changes in extremal dependence' because the method's stationarity assumption is unverified there; marginal scale changes during seizures could produce the same detector spikes.\n\nWhat the paper does well: the algorithm is clearly specified, computationally efficient (O(M n log G)), and reproducible – R package on GitHub. Simulation Scenario 2 (Student-t vs Gaussian copula with same correlation) is a genuine case where general-purpose nonparametric methods like E-divisive fail and MOPED succeeds. The permutation testing for thresholds is standard and not circular. The authors are also honest that there is no asymptotic theory; they flag it explicitly in Section 5.\n\nWhere the soft spots are. The biggest is the application. The model assumes stationary margins (Section 1), but the EEG preprocessing applies one global rank transform to the whole series. If seizure episodes change marginal variance or tail heaviness, the radial exceedances after ranking are not identically distributed, and the TPDM estimator can change even with constant copula. The abstract's claim about identifying 'structural changes in extremal dependence' in seizures is therefore stronger than the evidence supports. A supplementary analysis – testing homogeneity of marginal tails across windows, or using local marginal transforms – would settle this. It is a moderate concern, not a fatal one, because the methodology itself is about tail dependence under stationarity.\n\nSecond, the permutation null assumes serial independence. Subsampling every 256th point helps but does not remove all extremal dependence, and the threshold could then be too liberal. Also the simulation tables report averages without standard errors; a quick Monte Carlo standard error would help judge differences between MOPED and E-divisive.\n\nWho this is for: people working on change points in extremes, and applied statisticians with high-dimensional tail dependence. It deserves a serious referee: the missing theory is an honest gap, not a hidden flaw, and the method is novel and usable. I would send it to peer review, with a request to soften the EEG interpretation or add a check on marginal stationarity.","headline":"MOPED is a genuinely new and well-tested method for detecting changes in extremal dependence; the EEG conclusions are less secure because marginal changes can masquerade as dependence changes.","tokens_in":24666,"tokens_out":1881,"would_cite":true,"duration_ms":23040,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G32","62G10","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"MOPED uses a moving sum over tail-dependence matrices to find change points in multivariate extremes.","keywords":["change point detection","multivariate regular variation","tail dependence","tail pairwise dependence matrix","moving sum","electroencephalogram","seizure detection","extreme value theory"],"falsifier":"Simulate a long series with a constant TPDM but strong serial dependence in the extremes (e.g., a tail-dependent Markov chain), run MOPED with the nominal alpha permutation threshold, and count the fraction of runs with at least one detected change point. If that fraction exceeds alpha, the permutation null is not controlling the error rate under serial dependence.","tokens_in":23855,"feed_emoji":"🧠","tokens_out":5356,"duration_ms":58603,"temperature":0.7,"pith_summary":"The paper proposes MOPED, a nonparametric change-point detector for the dependence structure of multivariate extremes. It works by sliding a window along the time series, estimating the tail pairwise dependence matrix (TPDM) on each side of every candidate location, and flagging locations where the two local estimates differ enough to be unlikely under permutation. The method is designed to catch structural breaks in tail dependence even when the ordinary correlation structure is unchanged, a case that generic nonparametric detectors miss. The authors argue this makes tail-dependence change-point detection feasible in higher dimensions than existing extreme-value tests allow, and they demonstrate the payoff by locating seizure-related breaks in neonatal EEG recordings.","feed_headline":"A moving-sum scan locates shifts in extreme-value dependence","feed_subtitle":"Method catches structural changes in tail dependence that ordinary correlation-based detectors miss, with EEG seizure breaks as a test case.","key_machinery":"The tail pairwise dependence matrix (TPDM), the extreme-value analogue of a covariance matrix: entry (i,j) measures the asymptotic dependence between components i and j. MOPED's detector is a moving-sum (MOSUM) statistic that, at each time t and bandwidth G, subtracts the TPDM estimate from the G observations to the right of t from the estimate from the G observations to the left, and takes the Frobenius norm of the difference. Local maxima of this detector that exceed a permutation-calibrated threshold are declared change points. The multiscale, multi-threshold variant runs the detector over several bandwidths and radial thresholds and merges the resulting change-point sets with a bottom-up","core_discovery":"The central claim is that changes in the extremal dependence of a multivariate regularly varying time series can be detected and localized by a moving-sum statistic on TPDM estimates. Formally, the paper models the series as piecewise segments with segment-specific TPDMs, defines a detector D(G,t) that compares Frobenius norms of TPDM estimates in windows left and right of t, and declares change points at significant local maxima. A permutation procedure calibrates the threshold, and a multiscale multi-threshold variant pools estimates over bandwidths and exceedance thresholds. In simulations, the method identifies changes in tail dependence class that E-divisive misses, and in the EEG appli","pith_inferences":["Because the detector is built on the TPDM, MOPED could be pointed at any pairwise extremal summary by swapping the local estimator, yielding detectors tailored to asymptotic independence as well as asymptotic dependence.","A natural stress test of the permutation calibration would be to run MOPED on series with strong serial dependence but constant TPDM; the paper's EEG preprocessing subsamples every 256th point, suggesting residual dependence is expected to inflate false positives.","The bottom-up merging rule accepts all estimates from the finest bandwidth, so MMMOPED's false-positive propensity is likely concentrated at the smallest G; a stability check across permutations could rank change points by reproducibility.","In higher dimensions, the Frobenius norm averages evidence over all pairs; a weighted or pairwise-max version might localize which channels change, which could sharpen seizure-onset detection in EEG applications."],"forward_implications":["If correct, MOPED gives a nonparametric, multiple-change-point detector for tail dependence that works in dimensions where existing extreme-value tests are computationally prohibitive.","It can detect breaks in extremal dependence class (asymptotic dependence versus asymptotic independence) even when the Gaussian correlation structure stays constant, a case general-purpose detectors miss.","The method provides interpretable change points: each flagged break comes with a before/after TPDM, so practitioners can see which variable pairs changed.","In EEG monitoring, change points bracketing annotated seizures suggest tail-dependence changes are a usable automatic seizure signal.","The multiscale, multi-threshold variant reduces sensitivity to the classical threshold choice, though it is less conservative and can return spurious points."],"supporting_citations":[{"why":"Defines the TPDM and the empirical estimator that MOPED's detector is built on.","marker":"Cooley and Thibaud (2019)"},{"why":"Supplies the MOSUM detector framework and the local-maximizer criterion used for change-point estimation.","marker":"Eichinger and Kirch (2018)"},{"why":"Provides the E-divisive baseline that MOPED is compared against and the permutation idea for threshold selection.","marker":"Matteson and James (2014)"},{"why":"The parametric bivariate likelihood-ratio competitor in the simulation study.","marker":"Hazra and Bose (2025)"},{"why":"Contributes the bottom-up merging rule used to pool results across thresholds and bandwidths.","marker":"Messer et al. (2014)"},{"why":"The rolling sorted-values update that gives the O(n log G) computational complexity of the detector.","marker":"Vanegas et al. (2022)"},{"why":"Motivates the multiscale merging approach and supplies the lag-choice comparison.","marker":"McGonigle and Cho (2025)"},{"why":"Source of the neonatal EEG recordings and expert seizure annotations used in the application.","marker":"Stevenson et al. (2019)"}],"fun_headline_variants":["Moving-sum scan spots change points in tail dependence","New method detects shifts in extreme-value dependence","MOPED catches tail-dependence breaks correlation misses","A multiscale scan for change points in extremal dependence","Detecting structural breaks in multivariate extreme tails"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The permutation threshold is only valid if the observations are serially independent and the marginal distributions are stationary; otherwise the test can flag changes that are not actually changes in tail dependence, or miss true ones.","fun_headline_variants_meta":{"raw":{"variants":["Moving-sum scan spots change points in tail dependence","New method detects shifts in extreme-value dependence","MOPED catches tail-dependence breaks correlation misses","A multiscale scan for change points in extremal dependence","Detecting structural breaks in multivariate extreme tails"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000748,"raw_usage":{"total_tokens":3165,"prompt_tokens":737,"completion_tokens":2428,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":2365}},"tokens_in":481,"tokens_out":2428,"duration_ms":22083,"temperature":1.0,"reasoning_tokens":2365,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:26:28.081468+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a long series with a constant TPDM but strong serial dependence in the extremes (e.g., a tail-dependent Markov chain), run MOPED with the nominal alpha permutation threshold, and count the fraction of runs with at least one detected change point. If that fraction exceeds alpha, the permutation null is not controlling the error rate under serial dependence.","supporting_citations":[{"cited_title":"and Thibaud, E","cited_arxiv_id":null,"evidence_quote":"Defines the TPDM and the empirical estimator that MOPED's detector is built on."},{"cited_title":"and Kirch, C","cited_arxiv_id":null,"evidence_quote":"Supplies the MOSUM detector framework and the local-maximizer criterion used for change-point estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the E-divisive baseline that MOPED is compared against and the permutation idea for threshold selection."},{"cited_title":"and Bose, S","cited_arxiv_id":null,"evidence_quote":"The parametric bivariate likelihood-ratio competitor in the simulation study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the bottom-up merging rule used to pool results across thresholds and bandwidths."},{"cited_title":"J., Behr, M., and Munk, A","cited_arxiv_id":null,"evidence_quote":"The rolling sorted-values update that gives the O(n log G) computational complexity of the detector."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the multiscale merging approach and supplies the lag-choice comparison."},{"cited_title":"J., Tapani, K., Lauronen, L., and Vanhatalo, S","cited_arxiv_id":null,"evidence_quote":"Source of the neonatal EEG recordings and expert seizure annotations used in the application."}],"review_version":1}