{"id":"92820d64-bb71-47da-9d45-03bc730fadf8","arxiv_id":"2411.13822","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"New sparse high-dimensional extreme quantile regression with a refined Hill estimator, with rates for the tail index and the extrapolated quantile, plus simulations and an insurance data application.","lead":"A three-step estimator is proposed for extreme conditional quantiles when the number of predictors grows with the sample size, combining penalized quantile regression with extreme-value extrapolation. The paper proves convergence rates and reports better prediction accuracy than existing fixed-dimensional and central-quantile methods in simulations and an auto-insurance example.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The common-tail-index assumption (C5) is load-bearing: the extrapolation step reuses a single gamma for all x, and if gamma varies with covariates the claimed relative-error rate fails.","rationale":"Agree with the reader that Condition C5 is the weakest link. The three-step construction is clever and the rates in Theorems 1-3 are internally plausible, but the refined Hill step is a single-EVI estimator: it estimates γ at one covariate point and then assumes that index applies everywhere. That is precisely why C5 is stated, and it is also the reason the theorem is fragile. The proposal is a conditional result, and the paper is transparent about the condition in Remark 3, so this does not expose an internal inconsistency; it narrows the scope of the central claim. Because the reader already chose CONDITIONAL, and the missing supplement plus the impractical k rule further weaken verification, I recommend keeping the verdict unchanged rather than moving to accept or reject. The empirical demonstrations are useful but do not probe the assumption that would invalidate the theorem.","tokens_in":16613,"tokens_out":27839,"duration_ms":256914,"concrete_test":"Simulate from a process that satisfies the tail linear quantile model (1) but violates C5: let X∈{0,1} equally, and set QY(τ|X)=(1−X)(1−τ)^{-1}+X(1−τ)^{-2} for τ≥τln, so γ(0)=1 and γ(1)=2. Generate Y by inverse transform. Apply HEQR with the paper's recommended k rule, estimating γ at x̄≈0.5 (where γ=2), and compute relative error at τn=1−1/n at X=0 (true γ=1) for n=10^4,10^5,10^6. If the relative error fails to decrease (it should grow roughly like k), the common-EVI assumption is indispensable and Theorem 3 cannot hold without it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central consistency result (Theorem 3) is only as strong as Condition C5, which requires every conditional distribution Y|X to have the same extreme value index γ and the same tail shape up to a multiplicative factor K(Z) that is uniformly approximated. This assumption is what allows the refined Hill estimator to compute γ once at x̄ and then extrapolate to every covariate vector x. If the true tail index varies with X, the extrapolation factor ((1−τ0n)/(1−τn))^{Rath} is raised to the wrong exponent for most x, producing a relative bias of order ((1−τ0n)/(1−τn))^{γ(x)−γ(x̄)}−1, which diverges as τn→1 and completely dominates the Op(log[k/{n(1−τn)}](ns/k)sqrt(log(p∨n)/n)) rate in Theorem 3. The paper states this assumption in the missing supplementary material and gives no diagnostic or sensitivity analysis for it; Example 1 is a location-scale model where γ is constant by construction. Thus the announced consistency result applies only to a homothetic-tail family, a much narrower scope than the paper's title and abstract suggest.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-step estimator for extreme conditional quantiles in high-dimensional linear quantile regression with heavy tails: l1-penalized quantile regression at intermediate levels, a refined Hill estimator based on a fixed number of intermediate quantile fits, and Weissman-type extrapolation from an intermediate quantile to the extreme level. Theorems 1–3 state uniform error bounds for the intermediate quantile estimator, a convergence rate for the refined Hill estimator, and a relative-error rate for the extreme quantile estimator, all under conditions C1–C6 that are mostly deferred to a supplementary file. The paper also reports simulations and an auto-insurance application, and it discusses practical tuning for k, λ, and the refined Hill weights.","tokens_in":16895,"tokens_out":10686,"duration_ms":95849,"significance":"If the results are correct, the paper makes a useful contribution: it is among the first to derive uniform rates for penalized quantile regression at intermediate levels and to combine them with a refined Hill estimator in a high-dimensional setting, and it explicitly quantifies how the decay of the conditional density in the tail inflates the estimation rate. The proposed method is simple to implement, and the simulation study indicates that it can outperform both the fixed-dimensional extremal method of Wang et al. (2012) and the high-dimensional central-quantile method of Belloni and Chernozhukov (2011). However, the practical significance is tempered by two issues: the main results rest on a strong homogeneous-tail-index assumption (Condition C5), and the simulation settings violate the theoretical lower bound on k, so the reported numerical success is not covered by the stated theory.","major_comments":[{"comment":"The conditions C1–C4 and C6 and all proofs of Theorems 1–3 are in a supplementary file that is not included with the manuscript. The main text states Theorem 1 and Theorem 2 as consequences of these omitted conditions plus Lemma 1 and Remark S.2, which are also only referenced. Without the supplement, the central consistency claims cannot be verified; for example, the rate in Theorem 1 depends on unstated conditions on the design, the sparsity of θ_r, and the tail density, and the proof of Theorem 2 relies on a second-order expansion that is not shown. The supplement must be made available to reviewers before the technical content can be assessed.","section":"Section 2.1, Theorems 1–3; Section 2.4"},{"comment":"The assumption of a single extreme value index γ for all covariate values is load-bearing: the refined Hill estimator is computed at one point x̄ and then used in the extrapolation factor for every x. If the conditional tail index varies with X, the relative error of \\Q_Y(τ_n|x) contains a term ((1-τ0n)/(1-τn))^{γ(x)-γ(x̄)}-1 that diverges as τ_n→1 and dominates the rate in Theorem 3. The paper's title and abstract suggest a general method, but the theoretical result applies only to a homothetic-tail family. The simulations in Section 3 use a location-scale model with constant γ by construction, so they provide no evidence on robustness to heterogeneous tail indices. The paper should either narrow its claims, add a diagnostic for the common-tail assumption, or provide a sensitivity analysis.","section":"Section 2.1, Condition C5; Section 2.3, Remark 3"},{"comment":"The rule of thumb k = ⌊c0 n^{0.5+δ1} (log p)^{0.5+δ2}⌋ with (c0,δ1,δ2)=(0.8,0.01,0.05) yields k ≈ 54 for n=1000, p≈32 and k ≈ 137 for n=5000, p≈71, while the lower bound stated in Example 1(II) is √(log p)√n ≈ 59 and ≈ 146, respectively. Thus the simulation settings do not satisfy the assumptions under which Theorems 2 and 3 are proven. Although asymptotically the n^{δ1} factor eventually makes the rule exceed the lower bound, the sample sizes used in the simulations are not large enough for this to occur. Since the paper's simulation claims are a central part of the demonstration, the tuning rule should be revised (e.g., a larger c0) or the discrepancy should be discussed explicitly.","section":"Section 2.4, 'Selection of k'; Example 1(II)"}],"minor_comments":[{"comment":"The theorems require λ ≍ √n log p / √(1-τ0n), but all numerical results use 10-fold cross-validation. The cross-validated λ is not shown to satisfy the theoretical scaling, so the finite-sample results are not covered by the proof. This is a secondary gap, but it would help to comment on how CV-based λ relates to the theoretical choice.","section":"Section 2.4, 'Selection of λ'"},{"comment":"The symbol τ_n is used in the introduction for both the intermediate and the extreme quantile level, which conflicts with the later distinction between τ0n and τn. Please clarify the notation in the first occurrence.","section":"Introduction"},{"comment":"The caption for EQR reads 'the method in W ang, Li and He (2012)' with a typo in 'Wang'.","section":"Tables 1 and 2"},{"comment":"Some standard errors for HQR are large relative to the MISE values (e.g., 0.33 when MISE is 0.44), so the claim of 'significantly less efficient' is not supported by a formal comparison. Consider reporting confidence intervals or a paired comparison.","section":"Table 3"},{"comment":"The sentence 'Suppose there exist a positive sequence d1n such that, K1(Z) ≍ d1n hold for all Z ∈ Z.' repeats part of Condition C5; please rephrase or remove.","section":"Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The main technical results appear plausible, but the omitted supplement is a serious obstacle to verification. If the supplement is provided and the simulation tuning rule is reconciled with the theory (or the discrepancy is discussed), the paper could be publishable. The homogeneous-tail-index assumption should be stated prominently in the abstract and introduction, and a sensitivity analysis for heterogeneous γ would strengthen the practical claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: this is a legitimate first step on high-dimensional extreme quantile regression, but the title oversells the scope. The actual result is for a model where every conditional tail has the same extreme value index; the method estimates that one gamma at the mean covariate and extrapolates it to all x. If gamma varies with covariates, the relative error rate in Theorem 3 does not hold. The authors put this in Condition C5 and explicitly state the common-EVI consequence in Remark 3, so it is not a hidden flaw—but it is a real restriction, and the abstract doesn't advertise it.\n\nWhat's new: uniform tail bounds for L1-penalized quantile regression at intermediate levels, a refined Hill estimator built on a fixed number of penalized quantile fits, and plug-in Weissman extrapolation. That combination is new in high dimensions. The rates are plausible, and the simulation study shows HEQR beating both EQR and HQR in the settings they try. The location-scale example is a clean illustration of when the conditions hold.\n\nSoft spots, in order of importance.\n\nFirst, the practical k rule violates the theorem's lower bound in the simulations. Their rule k approximately 0.8 n^0.51 (log p)^0.55 is asymptotically faster than sqrt(n log p), but at n=1000 and 5000 with s=2 it lands below the required threshold. So the simulation regime is not covered by the theory. This is a finite-sample constant mismatch that is common, but it should be acknowledged or the rule adjusted.\n\nSecond, the regularity conditions and all proofs are in a missing supplement. I can't verify C1-C6 from the main text. Referees need that supplement to do their job.\n\nThird, the auto-insurance comparison uses EQR as a baseline, and EQR's prediction error has median absolute deviation exactly 0.00 across all CV splits. That smells like a degenerate baseline, so the headline \"HEQR wins\" is weaker than it looks. The authors should diagnose it.\n\nOn the stress-test note about C5: it lands. The common-tail-index assumption is load-bearing, and the consistency theorem really only covers a homothetic-tail family. The authors don't hide it, but the title and abstract overclaim relative to that scope.\n\nWho this is for: people working on extreme conditional quantile methods, and anyone applying high-dimensional QR to tails with reason to believe the tail index is constant. It deserves serious refereeing; the theory is non-trivial and the method has clear practical value. I'd accept for peer review, with a request for the supplement and a careful check of the tuning-rule gap.","headline":"A genuinely new high-dimensional extreme quantile method with a load-bearing common-tail-index assumption; the practical tuning rule does not fall under the stated theory and the supplement is missing, but the core is worth refereeing.","tokens_in":17351,"tokens_out":4009,"would_cite":true,"duration_ms":39331,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G32","62G20","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Extreme conditional quantiles can be estimated consistently when the number of predictors grows with the sample size, via penalized intermediate quantile regression, a refined Hill tail-index step, and extreme-value extrapolation.","keywords":["high-dimensional quantile regression","extreme conditional quantile","extreme value index","refined Hill estimator","ℓ1-penalized estimation","tail linear model","heavy-tailed distributions","quantile extrapolation"],"falsifier":"Run the method on simulated data from a sparse location-scale model with a covariate-dependent tail index — for example $Y=X_1+(1+0.5X_1)\\varepsilon$ with $\\varepsilon\\sim t(\\nu_1)$ for $X_2\\le 0$ and $\\varepsilon\\sim t(\\nu_2)$ for $X_2>0$, $\\nu_1\\neq\\nu_2$, with $p$ growing like $n$ — and check whether the refined Hill estimator stabilizes around a single γ and whether the relative error at $\\tau_n=0.999$ converges at the claimed rate. This violates Condition C5's common-EVI requirement, so the relative error should fail to vanish.","tokens_in":16450,"feed_emoji":"📈","tokens_out":15379,"duration_ms":121528,"temperature":0.7,"pith_summary":"This paper tries to close a gap: classical extreme-quantile regression assumes a fixed, small set of predictors, while high-dimensional quantile regression only targets central quantiles. The authors propose a three-step estimator — penalized quantile regression at intermediate levels, a refined Hill estimator of the tail index, and extreme-value extrapolation — and prove that it consistently estimates the extreme conditional quantile $Q_Y(\\tau_n|x)$ even when the dimension $p$ grows with the sample size or exceeds it. If correct, the result makes far-tail estimation feasible in settings where the number of recorded risk factors is large relative to the number of extreme events, as in insurance claims. The paper demonstrates the improvement in simulations and an auto-insurance application, where it also selects interpretable tail-relevant covariates.","feed_headline":"Extreme quantiles now reachable as predictors grow with sample size","feed_subtitle":"A three-step estimator outdoes direct high-dimensional quantile methods at 99.9th-percentile tails.","key_machinery":"The load-bearing object is the tail linear quantile model $Q_Y(\\tau|X)=Z^\\top\\beta(\\tau)$ for $\\tau\\in[\\tau_{ln},1)$ with only $s$ active slopes. On it the paper stacks three mechanisms: an $\\ell_1$-penalized quantile regression at intermediate levels, with penalty scale $\\lambda\\asymp \\sqrt{n\\log p}/\\sqrt{1-\\tau_{0n}}$; a refined Hill estimator $\\hat\\gamma=\\big(\\sum_{j=1}^J \\phi(\\ell_j)\\log(1/\\ell_j)\\big)^{-1}\\sum_{j=1}^J \\phi(\\ell_j)\\log\\big(z^\\top\\hat\\beta(\\tau_j)/z^\\top\\hat\\beta(\\tau_1)\\big)$ that combines a fixed number of intermediate quantile lines; and the extrapolation $\\hat Q_Y(\\tau_n|x)=\\big((1-\\tau_{0n})/(1-\\tau_n)\\big)^{\\hat\\gamma}z^\\top\\hat\\beta(\\tau_{0n})$. The argument runs on Condition C5, which assumes an auxiliary linear response $Z^\\top\\theta_r$ such that the tail of $U=Y-Z^\\top\\theta_r$ is uniformly close to one heavy-tailed distribution $F_0$ with a single extreme value index $\\gamma$ and controlled second-order error; that assumption supplies the density-decay rate that replaces the usual central-quantile density lower bound and justifies using one EVI estimate at every covariate value.","core_discovery":"On its own terms, the paper's central claim is that a sparse tail linear quantile model $Q_Y(\\tau|X)=Z^\\top\\beta(\\tau)$ for $\\tau$ close to 1 makes arbitrarily extreme conditional quantiles estimable in high dimensions. The estimator first runs $\\ell_1$-penalized quantile regression at intermediate levels $\\tau_j=1-\\ell_j k/n$, then estimates the extreme value index $\\gamma$ with a weighted refined Hill estimator applied to the fitted quantile lines, and finally extrapolates through the tail-quantile relation $U_Y(tz|x)/U_Y(t|x)\\to z^\\gamma$. Theorem 3 states that the relative error of the extrapolated estimator is $O_p\\big(\\log[k/\\{n(1-\\tau_n)\\}]\\,(ns/k)\\sqrt{\\log(p\\vee n)/n}\\big)$, which vanishes under the paper's conditions; the rate also requires the intermediate level $k$ to exceed $\\sqrt n$ by a factor and stay below $n$. Supporting this, the paper proves a uniform bound for the penalized intermediate-quantile estimator whose main inflation term, $d_n^\\gamma(1-\\tau_{0n})^{-\\gamma-1}$, measures how fast the conditional density decays at intermediate quantiles.","pith_inferences":["Editorial extension: the common-γ assumption is the main boundary; a grouped or local version of the refined Hill step would extend the method to data where the tail index varies with covariates, at the cost of pooling fewer observations per group.","Editorial extension: the paper's rule-of-thumb choice $k=\\lfloor c_0 n^{0.5+\\delta_1}(\\log p)^{0.5+\\delta_2}\\rfloor$ is heuristic; a data-driven selector that minimizes a plug-in estimate of the relative-error bound in Theorem 3 is a natural next test.","Editorial extension: the same three-step logic should transfer to other tail functionals such as extreme expectiles or conditional tail expectations by replacing the quantile loss and the extrapolation relation, though the density-decay argument would need re-derivation."],"forward_implications":["If the central claim is right, extreme conditional quantiles become estimable when $p$ is comparable to or larger than $n$, where the fixed-dimension EQR method is inapplicable and direct high-dimensional quantile regression has large bias.","The relative-error rate makes the bias–variance tradeoff explicit: consistency requires a window with $k$ above $\\sqrt n$, $k=o(n)$, and $\\log[k/\\{n(1-\\tau_n)\\}](ns/k)\\sqrt{\\log(p\\vee n)/n}\\to 0$, so the intermediate level cannot be chosen as in the fixed-dimension regime.","Because the $\\ell_1$ penalty produces sparse coefficient paths at intermediate quantiles, the same framework performs tail-relevant variable selection, not just quantile estimation.","The refined Hill estimator has a rate that does not depend on $\\gamma$, so the method remains stable across heavier-tailed responses where direct quantile estimators show much larger errors."],"supporting_citations":[{"why":"Supplies the fixed-dimensional extreme conditional quantile framework and the unweighted Hill extrapolation that HEQR extends, and serves as the EQR baseline in the simulations.","marker":"Wang, Li and He (2012)"},{"why":"Provides the ℓ1-penalized quantile regression estimator and penalty-calibration theory at central quantiles from which the intermediate step is adapted, and serves as the HQR baseline.","marker":"Belloni and Chernozhukov (2011)"},{"why":"States the tail-quantile relation U(tz)/U(t)→z^γ and the maximum-domain-of-attraction framework underlying the extrapolation step.","marker":"de Haan and Ferreira (2006)"},{"why":"Introduces the Hill estimator of the extreme value index that the refined weighted estimator generalizes.","marker":"Hill (1975)"},{"why":"Develops extreme quantile regression under density-tail conditions and provides the fixed-dimension template for extrapolating intermediate quantiles to the extreme level.","marker":"Chernozhukov (2005)"},{"why":"Gives the restricted strong convexity framework for high-dimensional M-estimators that the paper must adapt because the conditional density vanishes in the tail.","marker":"Negahban et al. (2012)"},{"why":"Supplies the weighted refined Hill estimator and the recommended weight parameters (J,s,a) that the paper adopts.","marker":"He et al. (2022)"},{"why":"Gives the principle that the penalty λ should dominate the rescaled subgradient, which the paper uses to calibrate λ at intermediate quantiles.","marker":"Bickel et al. (2009)"}],"fun_headline_variants":["Three-step estimator nails extreme quantiles in high dimensions","High-dim extreme quantiles? Sparse tail model + extrapolation works","When p grows with n, extreme quantiles still estimable","Tail quantile regression goes high-dimensional with Hill extrapolation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, after subtracting an auxiliary linear component, the tail of the response at every covariate value is uniformly close to one heavy-tailed distribution with a single extreme value index γ; if the tail index in fact varies across covariate values, or the uniform approximation error is too large, the extrapolation claim in Theorem 3 does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Three-step estimator nails extreme quantiles in high dimensions","High-dim extreme quantiles? Sparse tail model + extrapolation works","When p grows with n, extreme quantiles still estimable","Tail quantile regression goes high-dimensional with Hill extrapolation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00019,"raw_usage":{"total_tokens":1332,"prompt_tokens":927,"completion_tokens":405,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":335}},"tokens_in":543,"tokens_out":405,"duration_ms":4529,"temperature":1.0,"reasoning_tokens":335,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:50:28.693172+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on simulated data from a sparse location-scale model with a covariate-dependent tail index — for example $Y=X_1+(1+0.5X_1)\\varepsilon$ with $\\varepsilon\\sim t(\\nu_1)$ for $X_2\\le 0$ and $\\varepsilon\\sim t(\\nu_2)$ for $X_2>0$, $\\nu_1\\neq\\nu_2$, with $p$ growing like $n$ — and check whether the refined Hill estimator stabilizes around a single γ and whether the relative error at $\\tau_n=0.999$ converges at the claimed rate. This violates Condition C5's common-EVI requirement, so the relative error should fail to vanish.","supporting_citations":[],"review_version":1}