{"id":"440da3ab-8a18-40cb-898e-8871e60627ae","arxiv_id":"2607.15000","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Using block-graph gauge functions, the paper fits the first geometric extremal graphical model to 10 river gauging stations, enabling single-model estimates of simultaneous flood probabilities.","lead":"This paper builds a statistical model for estimating the chance of river floods happening at many places at the same time, using the shape of extreme river flows and the river network's own structure. It makes the 'geometric extremes' method usable at 10 gauging stations at once, so flood risk can be estimated from one fit instead of many.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Peak-matching pre-processing induces dependence that the model extrapolates; the 1.8e-4 annual probability is not robust to this data-construction step.","rationale":"The reader identified the peak-matching pre-processing as the weakest assumption, and I agree. The geometric graphical model is the first statistical implementation of a novel theoretical framework, and the paper is careful about fitting, diagnostics, and uncertainty (block bootstrap, threshold sensitivity, code availability). However, the central applied claim — an annual joint flood probability corresponding to a return period over 5000 years — is a pure extrapolation from a gauge function estimated on 2,582 matched points. The matching procedure replaces each event with per-station maxima over a ±2-day window, which is known to increase dependence (the paper states this). Since the bankfull thresholds are beyond observed data, the probability is entirely determined by the tail shape and dependence of the matched dataset. If matching inflates dependence, the probability is biased upward. The reported CIs are conditional on the matched data and do not account for the choice of p or the matching procedure itself. The reader's verdict CONDITIONAL is appropriate: the methodological novelty is sound and the analysis is competent, but the headline flood-risk estimate — and the 'strong extrapolation performance' claim — needs validation against the raw data or alternative pre-processings. The concrete test I propose directly targets the load-bearing assumption and would settle whether the concern lands. If the estimate is stable across p and comparison methods, the concern would be resolved; if not, the claim should be weakened or qualified.","tokens_in":24030,"tokens_out":2703,"duration_ms":23852,"concrete_test":"Rerun the full fitting and extrapolation pipeline on the original 15,459 daily observations without peak-matching, using a standard conditioning-site or max-stable baseline (or a geometric model fitted to raw daily data with a declustering threshold), and compare the estimated annual probability of simultaneous exceedance of v3,v4,v5,v7 at stations (3),(4),(5),(7). Also run the matching with p=1 and p=3 and re-estimate the same probability; if the estimate shifts by more than a factor of 2, the reported 1.8e-4 is not robust to the preprocessing choice.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claim — the annual joint flood probability 1.8[0.8,2.4]·10^-4 at stations (3),(4),(5),(7) in §5.2.4 — inherits all its extremal dependence information from the matched dataset constructed in §2.2/App. A.1. Matching replaces each event by the max flow within ±2 days at every station, so a rain storm that generates moderately high flows at several sites on different days is recorded as a single multivariate point with those maxima. This mechanically increases dependence: the paper itself notes that the matched data show stronger dependence (Fig. 2 and §2.2). Because the geometric gauge is estimated from these matched points, the fitted probability of simultaneous bankfull exceedance reflects the matching-induced correlation structure, not necessarily the event-level dependence that would produce a real flood. The matching window p=2 is chosen by inspecting one 1994 event (§2.2), with no sensitivity analysis over p. Also, the bankfull thresholds (v=860,318,580,900 m3/s) are far above observed maxima: no data points exceed bankfull at any of these stations, so the estimate is a pure extrapolation from the fitted gauge. The 95% block-bootstrap CIs reported (1.3-3.9·10^-6 per event) condition on the matched dataset and do not propagate uncertainty from p, the matching procedure, or the threshold choices. The claimed 'strong extrapolation performance' rests on in-sample diagnostics (PP plots, χ_S(u)); no out-of-sample or baseline comparison demonstrates that the extrapolation is unbiased. The weakest load-bearing link is therefore the unexamined assumption that the matched maxima are the true joint flood events; if matching inflates dependence, every extrapolated probability is biased upward, and the 5000-year return period is an artifact of preprocessing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents the first statistical implementation of geometric extremal graphical models, applying them to daily river flows at ten gauging stations on a river network near Preston, UK. The gauge function of the geometric limit set is parameterised through a block graph representing flow-connected and flow-unconnected stations (Eq. 7). The model is fitted via a truncated gamma model with kernel-based radial thresholds, and a correction coefficient C_corr (Eq. 3) is introduced to calibrate marginal probabilities. The authors assess fit via PP plots, limit-set projections and chi_S(u) comparisons, and use the fitted model to estimate the annual probability of simultaneous bankfull exceedance at stations (3),(4),(5),(7), reported as 1.8[0.8,2.4]*10^-4. An extended exponential-Gaussian gauge is proposed to allow joint extremes at selected pairs.","tokens_in":24364,"tokens_out":6272,"duration_ms":66212,"significance":"The methodological advance is real: it is the first parametric statistical implementation of geometric graphical models in a dimension (10) that is beyond previous geometric fits, and it offers a single-model alternative to conditional-extremes approaches with their conditioning-site ambiguity. The paper is careful in several respects: data and code are provided, block-bootstrap uncertainty quantification is used throughout, and the diagnostics are standard and clearly reported. The central quantitative claim, however, rests on a data-preprocessing step whose effect on dependence is acknowledged and not quantified, and on in-sample calibration. If the requested sensitivity analyses confirm stability, the method would be a useful addition to the flood-risk toolbox; as it stands, the numerical headline should be treated as conditional on the matching and threshold choices.","major_comments":[{"comment":"The matched dataset is constructed by replacing each event with the maximum flow within ±p=2 days at each station, reducing 15,459 to 2,582 points. As the paper notes, this mechanically strengthens dependence (Fig. 2). Since no observation exceeds bankfull at stations (3),(4),(5),(7) (§5.2.4), the annual probability 1.8[0.8,2.4]·10^-4 is a pure extrapolation whose dependence structure comes entirely from the matching-induced correlation. The window p=2 is chosen from one 1994 event, with no sensitivity analysis over p. The reported block-bootstrap CIs condition on the matched dataset and ignore uncertainty from p, the matching procedure, and the mass-balance filter. A sensitivity analysis over p (and ideally a comparison to an un-matched analysis) is needed to support the central quantitative claim.","section":"§2.2 / App. A.1"},{"comment":"The claim of 'strong extrapolation performance' (Abstract, §1) rests on the transformed PP plot (Figure 7) and χ_S(u) comparisons (Figure 9), which are in-sample diagnostics: both the empirical and model-based curves are computed from the same matched dataset and the same fitted model. These diagnostics do not validate extrapolation to the bankfull exceedance region in §5.2.4, where thresholds exceed observed maxima at all four stations. The χ_S(u) comparison is informative about dependence at observed quantile levels but not about the far-tail probability. The paper should include an out-of-sample or held-out event validation (e.g., fitting to a training period and evaluating on a test period, or a simulation-based calibration of the extrapolated probability) or temper the claim accordingly.","section":"§5.2.2–5.2.3"},{"comment":"The correction coefficient C_corr is an ad-hoc average of marginal probability ratios, justified only by 'numerical experiments' that are not shown. The reported probabilities, including 1.8·10^-4, are multiplied by this coefficient; different choices of S (which dimensions are included) and the marginal estimation used in the denominator can change the estimates. The paper should report the value of C_corr for the flood probability computation, show how the result changes without it, and provide a sensitivity analysis over S and over the exponential vs fitted marginal distribution.","section":"§3.2.3, Eq. (3)"},{"comment":"The extended exponential-Gaussian model has a relatively flat likelihood and wide CIs for γ parameters (e.g., γ_{5,7}: 1.00 [0.67,1.97]) and the bivariate projection for pair (5,7) is wider than the data suggest (Fig. 10). The claim that the model can accommodate both simultaneous and non-simultaneous extremes is therefore only weakly identified by the data. This is not a fatal flaw, but the paper should explicitly discuss what information in the data identifies γ, and avoid presenting the extended model's ability to detect joint extremes as a demonstrated property of this dataset.","section":"§6, Table 2"}],"minor_comments":[{"comment":"There is a typo: 'NRF A athttps://nrfa.ceh.ac.uk/' should be 'NRF A at https://nrfa.ceh.ac.uk/'.","section":"Data and code availability"},{"comment":"The axis labels 'GDF .9', 'GDF .5', etc. should be 'GDF_9', 'GDF_5' for readability.","section":"Figure 2"},{"comment":"The bankfull values v_exp,i are given without units; please clarify that these are in units of the standard exponential margins, not m³/s.","section":"§5.2.4"},{"comment":"The tolerance intervals in the transformed PP plot assume independent data. The paper notes they would be wider for dependent data, but it would be helpful to quantify the impact using the block-bootstrap samples.","section":"§5.2.2"},{"comment":"Please state explicitly what happens when S is empty or when b_s=0 for some dimensions; currently the definition of S is informal.","section":"§3.2.3, Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for an applied statistics journal. The central methodological idea is sound and the implementation is careful, but the headline quantitative claim in §5.2.4 depends on the peak-matching pre-processing and on a heuristic correction coefficient; those need robustness analyses before publication. The reader's circularity concern about the theoretical references is not, in my view, a blocker: the graphical-model results are independent of the application. I would not reject, but the requested sensitivity work is essential."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nThis is the first paper that actually implements geometric extremal graphical models rather than just developing the theory, and that is the main thing to know about it. The translation from Papastathopoulos & Wadsworth's block-graph gauge functions into a fitted 10-dimensional model for the Preston river network is careful and mostly convincing. For my money the genuinely new bits are: the block-graph construction keyed to river-network structure, the correction coefficient C_corr in eq (3), and the exponential-Gaussian extension that lets the data decide whether adjacent pairs are asymptotically dependent. The diagnostics are honestly done — MLE with block-bootstrap CIs, transformed PP plots, χ_S(u) comparisons, and a sensible sensitivity analysis around the influential point for pair (6,7). Code and data are public, which is real credit.\n\nThe soft spots are real but not fatal, and they cluster around the claim of 'strong extrapolation performance' and the specific annual probability in §5.2.4. The biggest one is the peak-matching preprocessing. The matched dataset is the entire dependence input, and matching ±2-day maxima at every station will mechanically inflate dependence — the paper says so itself. The window p=2 is chosen from one 1994 event, and there is no sensitivity analysis over p or over the 23-point removal rule. The stress-test note is right that the 1.8e-4 annual probability conditions on the matched dataset and the bootstrap CIs do not propagate uncertainty from the matching step. That said, I would not call this a load-bearing flaw in the methodology: the matching procedure is standard in hydrology, and the paper's central contribution is the graphical model implementation, not the precise value of one extreme joint probability. But the authors do lean on that value as an illustration, so the reviewer should ask for sensitivity to p and to the matching choice, and for some out-of-sample or baseline comparison before accepting the extrapolation claim.\n\nC_corr is admittedly a heuristic — the average of marginal ratios justified by numerical experiments rather than theory. That is a minor issue because the paper does not claim it is more than a practical fix, but it deserves scrutiny. The data-dependent choices (τ=0.85, Gaussian gauge for all pairs, k=2.7) are all disclosed and each is motivated; they make the fit look good but do not seem fraudulent.\n\nWho should read this: hydrologists and extremal-dependence statisticians who want a working recipe for fitting geometric models in 10+ dimensions. It deserves a serious referee. I would send it to review, with a request that the authors add a sensitivity analysis for the matching window and temper the extrapolation claim.","headline":"First real implementation of geometric extremal graphical models; sound methods and honest diagnostics, but the headline flood probability rests on a peak-matching preprocessing step that is not stress-tested.","tokens_in":25005,"tokens_out":1304,"would_cite":true,"duration_ms":12727,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G32","60G70","62P12"],"pacs":[],"model":"deepseek-v4-flash","headline":"One fitted model — a gauge function built from the river network's block graph — estimates any joint flood probability at 10 stations, including simultaneous floods at four stations with a return period over 5,000 years.","keywords":["geometric extremes","gauge function","extremal graphical models","block graphs","flood risk estimation","joint exceedance probabilities","river flow extremes","truncated gamma model"],"falsifier":"Re-fit the dependence model on the original 15,459 daily observations under several alternative alignment rules — no matching at all, ±1-day and ±3-day windows, and lags based on river travel times — and compare the resulting gauge parameters and the estimated annual probability of simultaneous flooding at stations (3), (4), (5) and (7). If that probability moves by orders of magnitude across the choices, the central claim collapses; if it stays within the reported 0.8–2.4 × 10^-4 interval, the matching assumption is vindicated.","tokens_in":23822,"feed_emoji":"🌊","tokens_out":20655,"duration_ms":173260,"temperature":0.7,"pith_summary":"This paper aims to bring the geometric approach to multivariate extremes to a scale that matters for flood risk: river networks typically have 10–20+ gauging stations, whereas geometric models had previously been fitted only in up to four dimensions, or up to ten with less interpretable machine-learning methods. The central proposal is to build the gauge function — the boundary curve describing how extreme flows can co-occur — out of the river network's block-graph structure, yielding a parsimonious 12-parameter model for the 10-station Preston network. The paper argues that one such fitted model can estimate any joint flood probability on the network, avoiding the awkward choice of a conditioning site that the conditional-extremes approach forces. As an illustration, it estimates the annual probability of simultaneous flooding at four stations as 1.8 × 10^-4, a return period beyond 5,000 years. If the approach is right, a single fit replaces a whole family of site-specific models for flood risk assessment.","feed_headline":"One fitted geometric model maps joint flood risk across 10 sites","feed_subtitle":"One 12-parameter fit replaces site-by-site models and can price events as rare as one in 5,000 years.","key_machinery":"The gauge function g(x) — the 1-homogeneous boundary of the limit set G = {x : g(x) ≤ 1}, whose shape dictates which variables can be extreme together. The paper builds g from a block graph (a decomposable graph with single-vertex separators, matching a river network's branching) as g(x) = Σ_C g_C(x_C) − Σ_D x_D; for Preston, six bivariate Gaussian gauges plus one four-dimensional clique gauge, minus x_3, x_4, x_5 and 3x_7. The truncated-gamma model R | {W = w, R > r_0(w)} ~ truncated Gamma(α, g(w)) makes the gauge a likelihood. The correction coefficient C_corr = (1/|S|) Σ_s exp(−b_s)/Pr̂(X_s > b_s) rescales estimates to offset marginal bias. The gauge carries the argument: it encodes depen","core_discovery":"The paper's central claim is that geometric extremal graphical models can be made statistical. The gauge function g — the 1-homogeneous boundary function whose unit set is the limit set of the scaled extreme cloud — is assembled from the river network's block graph as a sum of clique gauges minus separator coordinates; for the Preston network this becomes six bivariate Gaussian gauges for adjacent flow-connected pairs plus one four-dimensional Gaussian gauge for the flow-unconnected clique. Fitted through the truncated-gamma model for radial exceedances above an angle-dependent threshold, this 12-parameter gauge reproduces the observed tail dependence, and with a new correction coefficient i","pith_inferences":["Beyond the paper: the same block-graph recipe should transfer to other river networks of 10–20+ stations, because the parameter count grows with the number of edges and cliques rather than with the square of the dimension; the flow-unconnected clique is the bottleneck, since each unconnected station added there costs more clique parameters.","Beyond the paper: the correction coefficient is not specific to flood risk — any geometric-extremes application (rainfall, sea levels, spatial fields) that simulates from a fitted gauge inherits the same marginal bias, so C_corr can be lifted directly into those settings.","Beyond the paper: the single-fit property suggests a queryable flood-risk surface — joint probabilities for any demanded set of sites and thresholds — but its far-tail calibration rests on the matched-event construction, so a natural stress test is to repeat the analysis with matching windows of ±1 and ±3 days and see whether the 5,000-year estimate moves.","Beyond the paper: the event-set generation procedure in Appendix D makes annual exceedance probabilities for any region computable in principle, but it needed a million simulated years to produce about 100 events at the four-station bankfull region, so a cheaper far-tail sampler would be needed to make flood event catalogues practical."],"forward_implications":["A single fitted model answers joint flood probability questions for any combination of stations, replacing the four conditioning-site models that the conditional-extremes framework would need for the same four-station event — and it can handle combinations (such as stations 1 and 2) that no single conditioning model covers.","The annual probability of simultaneous flooding at stations (3), (4), (5) and (7) is estimated at 1.8 × 10^-4 (95% CI 0.8–2.4 × 10^-4), corresponding to a return period over 5,000 years; the extended model puts it at 1.6 × 10^-4.","The correction coefficient improves the model-based joint-tail coefficient χ_S(u) — the probability that all stations in a set S exceed a high quantile together — moving it closer to empirical values and narrowing the uncertainty as u → 1.","The graphical construction keeps the parameter count at 12 for 10 stations (six edge gauges plus six clique parameters) instead of the 45 parameters of a full 10-dimensional Gaussian gauge, which is what makes networks of 10–20+ stations tractable.","The exponential-Gaussian extension lets the data decide whether adjacent flow-connected pairs can experience joint extremes, with γ ≥ 1 estimated for pairs (3,7), (4,7) and (5,7); the paper notes the extrapolation differences from the simpler model appear only very far in the tail."],"fun_headline_variants":["12 parameters, 10 sites, one joint flood-risk map","One 12-parameter gauge prices rare floods at 10 sites","Geometric graph model: 10 sites, 12 params, 5,000-year events","Single geometric fit maps flood odds across 10 stations","Flood risk in one 12-parameter river-network fit"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the ±2-day peak-matching window turns raw daily flow records into genuine joint flood events: the paper itself notes the matched data show stronger dependence than the original series, so if the window is wrongly specified the fitted gauge and every extrapolated probability — including the 1.8 × 10^-4 annual flood probability — are biased.","fun_headline_variants_meta":{"raw":{"variants":["12 parameters, 10 sites, one joint flood-risk map","One 12-parameter gauge prices rare floods at 10 sites","Geometric graph model: 10 sites, 12 params, 5,000-year events","Single geometric fit maps flood odds across 10 stations","Flood risk in one 12-parameter river-network fit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000365,"raw_usage":{"total_tokens":1793,"prompt_tokens":731,"completion_tokens":1062,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":970}},"tokens_in":475,"tokens_out":1062,"duration_ms":10177,"temperature":1.0,"reasoning_tokens":970,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T00:28:12.778052+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-fit the dependence model on the original 15,459 daily observations under several alternative alignment rules — no matching at all, ±1-day and ±3-day windows, and lags based on river travel times — and compare the resulting gauge parameters and the estimated annual probability of simultaneous flooding at stations (3), (4), (5) and (7). If that probability moves by orders of magnitude across the choices, the central claim collapses; if it stays within the reported 0.8–2.4 × 10^-4 interval, the matching assumption is vindicated.","supporting_citations":[],"review_version":1}