Factorized Tail Volatility Model: Augmenting Excess-over-Threshold Method for High-Dimensional Hevay-Tailed Data

Yanxi Hou; Yifan Hu

REVIEW 3 major objections 4 minor 2 references

Factorized Tail Volatility Model: Augmenting Excess-over-Threshold Method for High-Dimensional Hevay-Tailed Data

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper introduces the Factorized Tail Volatility Model (FTVM), which writes each heavy-tailed observation as a low-rank factor-driven scale times a heavy-tailed idiosyncratic shock, and proves that fitting the factorization at an…

desk verdict A credible factor-model rate result that deserves a referee, but the FTVM-EoT headline rates rest on an unproved uniform threshold-convergence assumption. read the letter →

arxiv 2506.00840 v1 pith:EUWGXZC4 submitted 2025-06-01 stat.ME

Yifan Hu , Yanxi Hou This is my paper

classification stat.ME MSC 62G3262H2562G20

keywords extremevalueanalysisexcess-over-thresholdfactormodelheavy-taileddatatailquantileestimationregressionhigh-dimensionalselection

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a way to estimate very high tail quantiles for a large panel of heavy-tailed observations when each series has its own volatility. The central object, the Factorized Tail Volatility Model (FTVM), writes every observation as a low-rank factor-driven scale times a heavy-tailed idiosyncratic shock, so that heteroscedasticity in the bulk and heterogeneity in the tail are captured by the same factorization. The paper shows that fitting this factorization by quantile regression at an intermediate tail level recovers latent factors and loadings up to rotation with error $O_p(\sqrt{(N+T)/k})$, and that the resulting intermediate and extreme tail quantiles converge with mean squared relative errors of order $(N+T)/k$, with an extra log-squared term for extreme levels. It then augments the classical excess-over-threshold idea: a central-quantile model provides the threshold, and the FTVM models the excess beyond it, giving a modular framework that connects central, intermediate, and extreme quantiles. If correct, this gives applied tail-risk analysis a way to borrow strength across many series while still respecting each series' own heavy tail.

What carries the argument

The load-bearing device is the multiplicative factorization of a single tail quantile into a factor-driven scale and a tail-equivalent idiosyncratic quantile function, $Y_{i,t}=l_{0i}^{\top}f_{0t}U_{i,t}(V_{i,t}^{-1})$, together with the assumption that all $U_{i,t}$ are tail-equivalent to one reference function $U$ with extreme value index $\gamma>0$. Because the pairwise products $l_{0i}^{\top}f_{0t}$ are bounded away from zero and the tail quantiles are equivalent, the $k$-th largest order statistic of the pooled data consistently estimates the reference tail quantile $U(NT/k)$, and the normalized quantile check loss at level $k/(NT)$ identifies the factor structure up to rotation. The same normalized-loss device carries through the excess-over-threshold construction: after subtracting a central-quantile threshold estimator, the adjusted order statistics and Hill estimator inherit the original asymptotics provided the threshold error is uniformly $o_p(k^{-1/2})U(NT/k)$.

What would settle it

Generate data from the FTVM with known factors and a heavy tail, then estimate extreme quantiles using FTVM-EoT with a deliberately poor threshold estimator whose scaled uniform error is $O_p(k^{-1/3})$ instead of $O_p(k^{-1/2})$, violating Assumption 4. If the extreme-quantile mean squared relative error still decays at the advertised rate, the condition is not necessary; if it fails to decay, the advertised rates are confirmed to hinge on Assumption 4.

Watch

Extended reading notes

Core claim

The paper's central claim is that tail quantiles of high-dimensional heavy-tailed data can be factorized as $Y_{i,t}=l_{0i}^{\top}f_{0t}\,U_{i,t}(V_{i,t}^{-1})$, where $l_{0i}^{\top}f_{0t}$ is a bounded low-rank factor structure and $U_{i,t}$ are idiosyncratic tail quantile functions all tail-equivalent to a common reference function $U$. Estimating the factors and loadings by minimizing the quantile check loss at the intermediate level $k/(NT)$ on data scaled by the pooled $k$-th order statistic recovers the latent factors and loadings up to rotation with error $O_p(\sqrt{(N+T)/k})$, and yields intermediate-quantile mean squared relative errors of order $(N+T)/k$. For extreme quantile levels $p_{N,T}=o(k/(NT))$, extrapolation using the Hill estimator gives mean squared relative errors of order $(N+T)/k \vee \log^2(k/(NTp_{N,T}))/k$. The paper further claims that these properties transfer to the FTVM-EoT framework, where a central-quantile threshold model is subtracted first and the FTVM is fit to the excesses, provided the threshold estimator satisfies a uniform error condition.

Load-bearing premise

The entire excess-over-threshold extension rests on a condition the paper assumes rather than proves: the central-quantile threshold model must be uniformly accurate across all units and times, with its largest error after scaling by the pooled tail quantile shrinking faster than the reciprocal of the square root of $k$, where $k$ is the number of tail observations used.

Editorial extensions

If this is right

At intermediate tail quantile levels, the FTVM gives a factor-model estimate whose mean squared relative error is $O_p((N+T)/k)$, so the estimator improves as the panel grows provided $k$ grows faster than $N+T$.
At extreme quantile levels, extrapolation through the Hill estimator is formally justified with mean squared relative error $O_p((N+T)/k \vee \log^2(k/(NTp_{N,T}))/k)$.
Under the paper's uniform threshold-error condition, any central-quantile estimator can be plugged into the FTVM-EoT algorithm, and factor-number consistency, Kolmogorov-Smirnov validation, and quantile rates transfer to the enhanced estimator.
The degenerate model with zero factors is nested, and the proposed hypothesis test determines whether factor heterogeneity is strong enough to justify the FTVM over a simple pooled tail quantile estimate.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper leaves implicit is a data-driven rule for choosing $k$ by balancing the $(N+T)/k$ rate against the log-squared extreme-quantile term; the paper recommends a Hill plot but does not formalize the choice.
The modularity of the FTVM-EoT algorithm suggests that any central-quantile estimator with a proven uniform error bound could be upgraded to extreme-quantile estimation, but the paper demonstrates the plug-in idea only for two specific central-quantile models.
The simulations suggest one can test Assumption 4 empirically by checking whether the adjusted Kolmogorov-Smirnov statistic and adjusted Hill estimator behave as predicted before trusting extreme-quantile extrapolations; this would provide a practical diagnostic for the framework.
Because the simulation model uses Rademacher signs, the framework could be applied separately to positive and negative excesses to capture two-sided tail risk in financial returns, a use the paper mentions only implicitly.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Plain English: the paper's core rate result for its factorized tail volatility model (FTVM) is a genuine, if narrow, contribution. The advertised FTVM-EoT extension, however, is gated by Assumption 4, a uniform convergence condition on the central-threshold estimator that the authors explicitly assume rather than prove. That is the thing to know.

What is new: the joint quantile-regression estimator of factor scale at an intermediate tail level, with MSRE rates (N+T)/k for intermediate quantiles and a log-term version for extreme quantiles. Theorem 1's rates are plausible, the proof strategy is standard, and the supplementary material contains the proofs. The KS-based validation test and the information criterion for factor number are sensible supporting tools. The simulations are honest: they show the FTVM pays off only when heterogeneity is strong, and they report the small-sample failures (DGP2) rather than hiding them.

Soft spots, in proportion. The load-bearing one is Assumption 4. Proposition 3 and Corollary 1 need the max over i,t of the threshold-estimation error divided by U(NT/k) to be O_p(B_{N,T} k^{-1/2}). The paper says, in Section 5, 'we do not analyze the convergence of \hat H_{\tau*}, but instead assume its convergence as a condition.' The cited central-quantile rates from Ando & Bai (2020) and Chen et al. (2021) are pointwise or aggregate, not uniform over all i,t, so they do not obviously deliver (5.4). And for very heavy tails (lambda=1 in DGP4), the QFM density assumption breaks down. This is not a hidden defect—the authors are candid—but it means the headline rates for QFM-FTVM and QRIFE-FTVM are conditional on an unverified condition, which is a real gap between title promise and delivered theorems. The novelty claim that the idiosyncratic-effects issue 'has not been addressed in the literature' overstates prior work: Einmahl et al. (2014) and Hou et al. (2024) already handle heteroscedastic extremes in high-dimensional panels; the marriage is new, the components are not. The constraint bounds M and m are tuning parameters with material impact (Table 3), and the paper offers little guidance beyond defaults. No real-data application, no released code.

Who it is for: a methodological reader in econometrics or quantitative finance who wants factor-model parsimony and EVT tail extrapolation in one framework will take something from the rates and the algorithm. It deserves a serious referee. The referee's first question should be whether Assumption 4 can be proven for a concrete threshold model, or whether the paper should be reframed as conditional on that assumption. I would not desk-reject; I would send it out, and a revision should close or clearly foreground the Assumption 4 gap and add a real-data section.

Referee Report

3 major / 4 minor

Summary. The paper introduces the Factorized Tail Volatility Model (FTVM) for high-dimensional heavy-tailed panel data, in which tail quantiles are parameterized as l_i' f_t U_{i,t}(tau^{-1}) with a low-rank volatility factor component and a heteroscedastic tail-quantile component. The authors estimate factors and loadings by quantile check loss at an intermediate tail level, establish rates for factor/loading estimation and for intermediate and extreme tail quantiles, propose a Kolmogorov-Smirnov validation procedure and an information criterion for factor number selection, and then integrate FTVM with central quantile models (QFM and QRIFE) through an excess-over-threshold construction called FTVM-EoT. The main theoretical claims are Theorem 1 for FTVM, Theorem 2 for the factor-selection criterion, Proposition 3 for the adjusted sequence under Assumption 4, and Corollary 1 transferring the rates to FTVM-EoT.

Significance. If the claimed results hold, the paper offers a useful extension of the classical excess-over-threshold method to a high-dimensional factor structure, with explicit rates for intermediate and extreme quantiles, a validation tool, and a model-selection criterion. The FTVM-EoT construction is conceptually appealing because it separates a central-level location component from a tail-scale factor component, analogous to mean-volatility decompositions. I do not see an obvious flaw in the core FTVM estimation strategy: the rates in Theorem 1 are plausible, the check-loss formulation in (3.1) is natural, and the paper is honest about several finite-sample limitations. However, the headline FTVM-EoT results are conditional on Assumption 4, which is not proved for the specific threshold models used in the simulations. The paper also contains a sign/rate inconsistency in the information-criterion penalty that makes Theorem 2's conditions unsatisfied as written. These issues prevent acceptance in the current form but appear addressable.

major comments (3)

[Section 4, Eq. (4.3) and Theorem 2] The FTVM-EoT rates in Proposition 3 and Corollary 1 are gated by Assumption 4, which requires a uniform max error bound of the form max_{i,t} |\hat H_{tau*}(I_{i,t}) - H_{0,tau*}(I_{i,t})| / U(NT/k) = O_p(B_{N,T} k^{-1/2}) with B_{N,T}->0. The paper explicitly states in Section 5: "we do not analyze the convergence of \hat H_{tau*}, but instead assume its convergence as a condition." The cited central-quantile rates from Ando & Bai (2020) and Chen et al. (2021) are not shown to deliver this uniform-max bound, and the paper's own DGP4/DGP5 simulations with lambda=1 show QFM and QRIFE degrading in a heavy-tail regime. Thus the advertised QFM-FTVM and QRIFE-FTVM rates are conditional on a condition that is never verified. Please either prove Assumption 4 for these threshold models under the stated assumptions or explicitly re-frame the FTVM-EoT results as conditional on this unverified high-level condition.
[Section 5, Proposition 3 proof] The penalty defined in (4.3) has the factor log(k/(N+T)). Since k/(N+T)->0, this logarithm is negative, so the penalty P_{N,T} is negative and P_{N,T}(k/(N+T)) tends to -infinity, contradicting Theorem 2's condition that P_{N,T}(k/(N+T))->infinity. The sign of the penalty is load-bearing: with a negative penalty, the information criterion in (4.2) can only reward larger l and cannot be consistent. The formula and the surrounding text should presumably use log((N+T)/k); the inconsistency also propagates into Remark 4 and the simulation description of the penalty with c=10.
[Section 6.2.2, Table 3] Even granting Assumption 4, the proof of Proposition 3 is incomplete because it treats the estimated threshold \hat H_{tau*} as a deterministic shift: it checks that the adjusted sequence has the same tail-equivalence property as the original FTVM. But \hat H_{tau*} is estimated from the full NT sample, so the adjusted residuals Y_{i,t} - \hat H_{tau*}(I_{i,t}) are no longer independent across i and t. The order-statistic results and the KS empirical-process convergence require a coupling or stochastic-equicontinuity argument to control this induced dependence; the current proof does not supply one.

minor comments (4)

[Section 6.2.2, Table 3] The performance of the constrained estimator in (3.1) is highly sensitive to the tuning parameter M. For example, Table 3 shows the MSRE rising from 255 to 8919 (times 10^{-3}) as M increases from 1.3 to 32 at (N,T)=(50,50), lambda=1, k=0.1NT. The theoretical rates in Theorem 1 do not display a dependence on M, so the paper would benefit from explicit guidance on choosing m and M, or from a sensitivity analysis that reconciles the finite-sample overfitting shown in Figure 1 with the asymptotic theory.
[Section 5, KSadj definition] The adjusted KS statistic KSadj is defined immediately after introducing \hat U^adj(NT/k), but the displayed formula uses \hat U(NT/k) in the indicator. If the unadjusted threshold is intentional, please explain why; otherwise the notation should be \hat U^adj(NT/k).
[Section 6.2.3, near end] The text states that for (N,T)=(50,50) and k=0.05NT the ratio k/(N+T) equals 2.5. In fact k=0.05*2500=125 and N+T=100, so k/(N+T)=1.25; the value 2.5 corresponds to k=0.1NT. The sentence's explanation of when Theorem 1's conditions are violated should be corrected.
[Throughout] There are numerous typos and minor errors that should be fixed before publication: "Ecess" in the abstract, "Hevay" in the title, "assummption" and "componet" in Section 2, "varaible" in Example 2, "estiamte" and "intermedaite" in Section 6.3, and "anf" in the caption of Table 3.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the FTVM-EoT rates are explicitly conditional on an unproved Assumption 4, which is a correctness gap rather than a circular reduction.

full rationale

The derivation chain is self-contained. The FTVM estimators in (3.1) are standard quantile-regression M-estimators fitted at the intermediate level k/NT; reporting the MSRE at that same level is in-sample estimation, not a renamed input or a forced prediction. The extreme-quantile estimator uses the Hill extrapolation from an intermediate threshold, and the target level p_{N,T} is never used in fitting the factor/loading parameters, so the extreme prediction is not equivalent to its inputs by construction. Proposition 1's use of the pooled order statistic as the reference U is justified by Assumptions 2-3 together with the scale normalization (2.3), which is an identification condition rather than a tautological definition. The FTVM-EoT results (Proposition 3 and Corollary 1) are explicitly conditional on Assumption 4; the paper states in Section 5: "we do not analyze the convergence of \hat H_{tau*}, but instead assume its convergence as a condition." This is an unproved sufficient condition, and therefore a correctness/robustness gap for the QFM/QRIFE implementations, but assuming a condition is not a circular step. The only author-overlapping citation, Hou et al. (2024), is background in the introduction and is not used to justify any theorem; it is an externally published paper and is not load-bearing. No step equates a fitted parameter with the target prediction by construction.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The framework imports standard factor-model identification and EVT second-order conditions. The clearest extra assumptions are the normalization (2.3) forcing the pooled average scedasis to one, and Assumption 4 postulating uniform convergence of the threshold model. No new physical entities are introduced. The user-chosen tuning parameters M, m, k, and c materially affect finite-sample performance.

free parameters (5)

M (upper bound in constraint (3.1)) = 1.6 in main simulations; varies from 1 to 32 in sensitivity analysis
Bounds the estimated factorized volatility l_i' f_t. No data-driven selection rule is given, and MSRE strongly depends on it (Table 3).
m (lower bound in constraint (3.1)) = 0.1 or 0.01 in simulations
Keeps volatilities bounded away from zero, as required by Assumption 1, and affects the feasible set of the optimization.
k (intermediate order statistic level) = 0.1 NT and 0.05 NT in simulations
User-selected number of tail observations. All estimators and theoretical rates depend on k; the paper recommends a Hill plot rather than giving a formal selection rule.
c (penalty constant in information criterion (4.3)) = 10 in simulations
Scales the factor-number penalty in the information criterion. No systematic selection rule is provided.
tau* (central threshold quantile in FTVM-EoT) = 0.5 in simulations
The split between the central threshold model and the tail model is user-chosen; performance at extreme quantiles may depend on it.

assumptions (5)

domain assumption Identification constraints: loadings and factors lie in compact sets, min_{i,t} l_i' f_t >= m > 0, N^{-1} L L' tends to a diagonal matrix, and T^{-1} F F' = I_r.
Assumption 1, Section 2. Needed for factor identification and non-degenerate quantiles.
domain assumption Tail-equivalence of all U_{i,t} to a common U with the second-order condition in equation (2.2).
Assumption 2 and equation (2.2), Section 2. This is a standard second-order EVT condition imported from de Haan and Ferreira.
ad hoc to paper Normalization sqrt{k} | (NT)^{-1} sum_{i,t} (l_i' f_t)^{1/gamma} - 1 | tends to zero.
Equation (2.3), Section 2. Makes U the unconditional pooled tail quantile; without this condition, the pooled order statistic is not the reference quantile.
domain assumption Conditional on factors and loadings, the epsilon_{i,t} are independent with uniform V_{i,t}.
Model definition in Section 1. Excludes tail dependence beyond the factor structure.
ad hoc to paper Assumption 4: the central threshold estimator hat H_{tau*} satisfies max_{i,t} |hat H_{tau*}(I_{i,t}) - H_{0,tau*}(I_{i,t})| / U(NT/k) = O_p(B_{N,T} k^{-1/2}).
Section 5. The paper states it is not proved for QFM or QRIFE, yet Proposition 3 and Corollary 1 depend on it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Factorized Tail Volatility Model: Augmenting Excess-over-Threshold Method for High-Dimensional Hevay-Tailed Data." pith.science (2026). https://pith.science/paper/EUWGXZC4

@misc{pith2026250600840,
  author       = {Pith},
  title        = {Pith review of: Factorized Tail Volatility Model: Augmenting Excess-over-Threshold Method for High-Dimensional Hevay-Tailed Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUWGXZC4}},
  note         = {Machine review of arXiv:2506.00840}
}

read the original abstract

Ecess-over-Threshold method is a crucial technique in extreme value analysis, which approximately models larger observations over a threshold using a Generalized Pareto Distribution. This paper presents a comprehensive framework for analyzing tail risk in high-dimensional data by introducing the Factorized Tail Volatility Model (FTVM) and integrating it with central quantile models through the EoT method. This integrated framework is termed the FTVM-EoT method. In this framework, a quantile-related high-dimensional data model is employed to select an appropriate threshold at the central quantile for the EoT method, while the FTVM captures heteroscedastic tail volatility by decomposing tail quantiles into a low-rank linear factor structure and a heavy-tailed idiosyncratic component. The FTVM-EoT method is highly flexible, allowing for the joint modeling of central, intermediate, and extreme quantiles of high-dimensional data, thereby providing a holistic approach to tail risk analysis. In addition, we develop an iterative estimation algorithm for the FTVM-EoT method and establish the asymptotic properties of the estimators for latent factors, loadings, intermediate quantiles, and extreme quantiles. A validation procedure is introduced, and an information criterion is proposed for optimal factor selection. Simulation studies demonstrate that the FTVM-EoT method consistently outperforms existing methods at intermediate and extreme quantiles.

Figures

Figures reproduced from arXiv: 2506.00840 by the authors.

**Figure 1.** Scatter plot of (log(Yi,t), log( ˆl ⊤ i,2 ˆft,2Uˆ(10))), where Yi,t is generated from DGP3 with λ = 1, (N, T) = (50, 50), and ˆl ⊤ i,2 , ˆf ⊤ t,2 are estimated by FTVM with r = 2, m = 0.1 and M = 32 The solid line represents the line y = x. Notably, the performance of the FTVM worsens when M becomes excessively large. For instance, when M = 32 with (N, T) = (50, 50), MSRE0.1 increases significantly to 8.919. A plaus… view at source ↗

**Figure 2.** Boxplot of the Hill estimator for different [PITH_FULL_IMAGE:figures/full_fig_p032_2.png] view at source ↗

**Figure 3.** For Yi,t generated from DGP4, we plot the quantiles of Yi,t at pN,T = 0.001(top-left) and pN,T = 0.0001(bottom-left), estimated quantiles by EoTM at pN,T = 0.001(top-middle) and pN,T = 0.0001(bottom-middle), and estimated quantiles by QFM at pN,T = 0.001(topright) and pN,T = 0.0001(bottom-right). We conduct several simulation experiments to evaluate the performance of the FTVM and the FTVM-EoT approach under variou… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

& Bai, J

Ando, T. & Bai, J. (2020), ‘Quantile co-movement in financial markets: A panel quantile model with unobserved heterogeneity’,Journal of the American Statistical Association 115(529), 266–279. Barigozzi, M. & Hallin, M. (2020), ‘Generalized dynamic factor models and volatilities: Consistency, rates, and prediction intervals’,Journal of Econometrics216(1), ...

work page 2020
[1356]

Einmahl, J. H. J., Haan, L. & Zhou, C. (2014), ‘Statistics of Heteroscedastic Extremes’, Journal of the Royal Statistical Society Series B: Statistical Methodology78(1), 31–51. Haan, L. & Ferreira, A. (2006),Extreme value theory: an introduction, Vol. 3, Springer. Hou, Y., Leng, X., Peng, L. & Zhou, Y. (2024), ‘Panel quantile regression for extreme risk’,...

work page 2014

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

& Bai, J

Ando, T. & Bai, J. (2020), ‘Quantile co-movement in financial markets: A panel quantile model with unobserved heterogeneity’,Journal of the American Statistical Association 115(529), 266–279. Barigozzi, M. & Hallin, M. (2020), ‘Generalized dynamic factor models and volatilities: Consistency, rates, and prediction intervals’,Journal of Econometrics216(1), ...

work page 2020

[2] [1356]

Einmahl, J. H. J., Haan, L. & Zhou, C. (2014), ‘Statistics of Heteroscedastic Extremes’, Journal of the Royal Statistical Society Series B: Statistical Methodology78(1), 31–51. Haan, L. & Ferreira, A. (2006),Extreme value theory: an introduction, Vol. 3, Springer. Hou, Y., Leng, X., Peng, L. & Zhou, Y. (2024), ‘Panel quantile regression for extreme risk’,...

work page 2014