pith. sign in

arxiv: 2606.28540 · v1 · pith:72W5KKTWnew · submitted 2026-06-26 · 📊 stat.ME

Choosing the threshold in extreme value analysis

L\'eo R. Belzile , Anthony C. Davison This is my paper

Pith reviewed 2026-06-30 01:07 UTC · model grok-4.3

classification 📊 stat.ME
keywords threshold selectionextreme value analysisgeneralized Pareto distributionsimulation studyHill estimatorgoodness-of-fit testsrainfall dataunivariate extremes
0
0 comments X

The pith

Simulations identify the strongest performers among more than 40 threshold selection methods for extreme value analysis.

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

The paper reviews more than 40 procedures for selecting thresholds when modeling exceedances over a large value in univariate extreme value analysis. It examines their underlying statistical properties, evaluates strengths and weaknesses, considers how they can be automated, and compares them via an extensive simulation study together with a long series of daily rainfall totals. A sympathetic reader would care because the threshold choice strongly affects the resulting parameter estimates and predictions, yet its uncertainty is commonly ignored in applications. If the simulation findings are correct, analysts obtain concrete guidance on which procedures to prefer in practice.

Core claim

Through a systematic review of the statistical foundations of more than 40 threshold selection procedures—including those based on Hill's estimator, visual diagnostics, goodness-of-fit tests, and extended generalized Pareto models—the paper identifies the most promising methods via an extensive simulation study and applies the comparison to daily rainfall data from Padova.

What carries the argument

Threshold selection procedures that determine the cutoff value above which exceedances are modeled by the generalized Pareto distribution.

If this is right

  • Certain procedures can be automated for routine use without requiring manual diagnostics.
  • Inference for extremes becomes more stable when the threshold is chosen by one of the top-ranked methods.
  • The uncertainty attached to the threshold can be incorporated into subsequent risk calculations more readily.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same simulation framework could be applied to test whether the top methods remain reliable under stronger serial dependence or non-stationary conditions.
  • Extension to multivariate extremes would require new criteria because joint threshold choice interacts with dependence modeling.
  • Bayesian formulations could treat the threshold as a random variable and compare posterior performance against the frequentist selectors identified here.

Load-bearing premise

The chosen simulation scenarios and the single Padova rainfall series adequately represent the range of real-world data and dependence structures that arise in applications.

What would settle it

A fresh simulation study that employs substantially different data-generating processes and produces a markedly different ranking of the procedures would falsify the identification of the most promising methods.

read the original abstract

One of the two dominant approaches for univariate extreme value analysis is to model exceedances above a large threshold, the choice of which has a large impact on inference and whose uncertainty is often subsequently ignored. In this article we review more than 40 threshold selection procedures, including semiparametric methods based on Hill's estimator, visual diagnostics, goodness-of-fit tests, and others based on extended generalized Pareto models. Starting with the statistical properties underlying the various proposals, we provide a critical assessment of their strengths and weaknesses, discuss how they might be automated and describe the results of an extensive simulation study used to identify the most promising procedures. The approaches are compared using a long time series of daily rainfall totals from Padova.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper reviews more than 40 threshold selection procedures for univariate extreme value analysis (including Hill-based, visual, GoF, and extended GPD methods), critically assesses their statistical properties and automation potential, and reports results from an extensive simulation study to identify the most promising procedures, with comparison also performed on a long daily rainfall series from Padova.

Significance. If the simulation-based ranking holds under representative conditions, the work would offer practitioners concrete guidance on a high-impact modeling choice whose uncertainty is frequently ignored, potentially improving reliability of tail inferences across applications.

major comments (2)
  1. [Simulation study] Simulation study section: the data-generating processes are not described as including temporal clustering, non-stationarity, or heavy-tail regimes with varying dependence; because the Padova rainfall validation series exhibits these features, the reported ranking of the >40 procedures may not generalize and the central comparative claim rests on an unverified assumption of representativeness.
  2. [Application to rainfall data] Validation on Padova series: the single real-data example is used to illustrate but no quantitative comparison of method performance (e.g., stability of parameter estimates or out-of-sample predictive scores) is reported against the simulation winners, weakening the link between simulation rankings and practical utility.
minor comments (2)
  1. The abstract states 'more than 40' methods but the main text should include an explicit table or appendix listing all reviewed procedures with their key references for reproducibility.
  2. Notation for threshold estimators (e.g., u_n, k) should be standardized across sections to avoid reader confusion when comparing semiparametric and model-based approaches.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address the two major comments below, agreeing where the critique is valid and outlining targeted revisions. Our responses focus on clarifying the scope of the work while strengthening the manuscript's transparency and practical relevance.

read point-by-point responses

  1. Referee: [Simulation study] Simulation study section: the data-generating processes are not described as including temporal clustering, non-stationarity, or heavy-tail regimes with varying dependence; because the Padova rainfall validation series exhibits these features, the reported ranking of the >40 procedures may not generalize and the central comparative claim rests on an unverified assumption of representativeness.

    Authors: We agree that the simulation study is conducted under the classical i.i.d. setting with standard GPD tails and does not incorporate temporal clustering, non-stationarity, or dependence structures. This choice was deliberate to isolate the performance of threshold selectors under the asymptotic conditions where most of the reviewed methods are theoretically justified. The Padova series is presented as a separate real-data illustration rather than a direct validation of the simulation rankings. In revision we will (i) explicitly document the simulation assumptions in a dedicated limitations subsection, (ii) add a brief discussion of how dependence and non-stationarity may affect the relative performance of the procedures, and (iii) caution readers that the reported ranking applies most directly to the i.i.d. case. We do not plan to rerun the full simulation suite with dependent or non-stationary DGPs, as that would constitute a substantially different study. revision: partial

  2. Referee: [Application to rainfall data] Validation on Padova series: the single real-data example is used to illustrate but no quantitative comparison of method performance (e.g., stability of parameter estimates or out-of-sample predictive scores) is reported against the simulation winners, weakening the link between simulation rankings and practical utility.

    Authors: The Padova analysis is intended as an illustrative case study rather than a formal validation exercise, precisely because the true threshold is unknown. We therefore did not compute stability or predictive metrics that would require additional modeling assumptions. We accept that this leaves a gap between the simulation results and practical guidance. In the revised manuscript we will add a short quantitative subsection that reports, for the top-ranked procedures, (a) the variability of the resulting GPD parameter estimates across bootstrap resamples of the series and (b) a simple out-of-sample check using a hold-out period to compare predictive performance of the fitted tails. These additions will be presented as supplementary evidence rather than definitive proof of superiority. revision: yes

Circularity Check

0 steps flagged

No circularity: review and simulation study with independent empirical basis

full rationale

The paper reviews over 40 existing threshold selection methods for extreme value analysis, provides critical assessment based on their statistical properties, and ranks them via an extensive simulation study plus application to one rainfall series. No mathematical derivation chain exists that reduces predictions or results to inputs by construction, self-definition, or self-citation load-bearing. The simulation outcomes and comparisons are generated independently from the reviewed methods and do not involve fitting parameters then relabeling them as predictions. This is a standard empirical review paper whose claims rest on external simulation design rather than internal circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review contains no identifiable free parameters, axioms, or invented entities; the simulation study itself would normally introduce modeling choices but none are stated here.

pith-pipeline@v0.9.1-grok · 5641 in / 1029 out tokens · 22524 ms · 2026-06-30T01:07:22.037708+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

210 extracted references · 122 canonical work pages

  1. [1]

    1928 , journal =

    Limiting forms of the frequency distributions of the largest or smallest member of a sample , author =. 1928 , journal =

    1928

  2. [2]

    1943 , journal =

    Sur la distribution limite du terme maximum d'une s\'erie al\'eatoire , author =. 1943 , journal =

    1943

  3. [3]

    1953 , journal =

    On the theory of order statistics , author =. 1953 , journal =. doi:10.1007/BF02127580 , isbn =

    work page doi:10.1007/bf02127580 1953

  4. [4]

    1954 , journal =

    La distribution de la plus grande de n valeurs , author =. 1954 , journal =

    1954

  5. [5]

    1955 , journal =

    The frequency distribution of the annual maximum (or minimum) values of meteorological elements , author =. 1955 , journal =

    1955

  6. [6]

    1958 , publisher =

    Statistics of Extremes , author =. 1958 , publisher =

    1958

  7. [7]

    A general definition of residuals (with

    Cox, David R and Snell, E Joyce , year =. A general definition of residuals (with. Journal of the Royal Statistical Society. Series B , publisher =

  8. [8]

    1970 , journal =

    A stochastic model for flood analysis , author =. 1970 , journal =. doi:10.1029/WR006i006p01641 , url =

    work page doi:10.1029/wr006i006p01641 1970

  9. [9]

    1971 , journal =

    Some problems of flood analysis , author =. 1971 , journal =. doi:10.1029/WR007i005p01144 , url =

    work page doi:10.1029/wr007i005p01144 1971

  10. [10]

    , TITLE =

    Residual life time at great age , author =. 1974 , journal =. doi:10.1214/aop/1176996548 , url =

    work page doi:10.1214/aop/1176996548 1974

  11. [11]

    1975 , journal =

    A simple general approach to inference about the tail of a distribution , author =. 1975 , journal =

    1975

  12. [12]

    1975 , journal =

    Statistical inference using extreme order statistics , author =. 1975 , journal =

    1975

  13. [13]

    1978 , journal =

    Estimation of a biometric function , author =. 1978 , journal =

    1978

  14. [14]

    1978 , journal =

    Estimation of parameters and large quantiles based on the k largest observations , author =. 1978 , journal =

    1978

  15. [15]

    1982 , journal =

    Maximum likelihood estimation of misspecified models , author =. 1982 , journal =. doi:10.2307/1912526 , url =

    work page doi:10.2307/1912526 1982

  16. [16]

    1983 , publisher =

    Extremes and Related Properties of Random Sequences and Processes , author =. 1983 , publisher =

    1983

  17. [17]

    1984 , booktitle =

    Modelling excesses over high thresholds, with an application , author =. 1984 , booktitle =. doi:10.1007/978-94-017-3069-3\_34 , url =

    work page doi:10.1007/978-94-017-3069-3 1984

  18. [18]

    1984 , journal =

    Testing whether the shape parameter is zero in the generalized extreme-value distribution , author =. 1984 , journal =

    1984

  19. [19]

    1984 , booktitle =

    Threshold methods for sample extremes , author =. 1984 , booktitle =

    1984

  20. [20]

    1985 , journal =

    Estimation of the generalized extreme-value distribution by the method of probability-weighted moments , author =. 1985 , journal =

    1985

  21. [21]

    1985 , journal =

    Adaptive estimates of parameters of regular variation , author =. 1985 , journal =. doi:10.1214/aos/1176346596 , url =

    work page doi:10.1214/aos/1176346596 1985

  22. [22]

    1985 , journal =

    Kernel estimates of the tail index of a distribution , author =. 1985 , journal =. doi:10.1214/aos/1176349656 , url =

    work page doi:10.1214/aos/1176349656 1985

  23. [23]

    1985 , journal =

    Maximum likelihood estimation in a class of nonregular cases , author =. 1985 , journal =. doi:10.1093/biomet/72.1.67 , url =

    work page doi:10.1093/biomet/72.1.67 1985

  24. [24]

    J. R.M. Hosking and J. R. Wallis , year =. Parameter and quantile estimation for the generalized. Technometrics , publisher =. doi:10.1080/00401706.1987.10488243 , url =

    work page doi:10.1080/00401706.1987.10488243 1987

  25. [25]

    1987 , publisher =

    Extreme values, regular variation, and point processes , author =. 1987 , publisher =

    1987

  26. [26]

    , author =

    Approximations in extreme value theory. , author =. 1987 , url =

    1987

  27. [27]

    1989 , journal =

    On the estimation of the extreme-value index and large quantile estimation , author =. 1989 , journal =. doi:10.1214/aos/1176347396 , url =

    work page doi:10.1214/aos/1176347396 1989

  28. [28]

    1989 , journal =

    A method for the statistical definition of extreme-value regions and their application to meteorological time series , author =. 1989 , journal =

    1989

  29. [29]

    Models for exceedances over high thresholds (with

    Davison, Anthony C and Smith, Richard L , year =. Models for exceedances over high thresholds (with. Journal of the Royal Statistical Society. Series B , publisher =

  30. [30]

    1990 , journal =

    Models for exceedances over high thresholds (with Discussion) , author =. 1990 , journal =. doi:10.1111/j.2517-6161.1990.tb01796.x , url =

    work page doi:10.1111/j.2517-6161.1990.tb01796.x 1990

  31. [31]

    1990 , journal =

    L -moments: Analysis and estimation of distributions using linear combinations of order statistics , author =. 1990 , journal =. doi:10.1111/j.2517-6161.1990.tb01775.x , url =

    work page doi:10.1111/j.2517-6161.1990.tb01775.x 1990

  32. [32]

    1990 , journal =

    Using the bootstrap to estimate mean squared error and select smoothing parameter in nonparametric problems , author =. 1990 , journal =. doi:10.1016/0047-259x(90)90080-2 , url =

    work page doi:10.1016/0047-259x(90)90080-2 1990

  33. [33]

    and Spurr, B.D

    Sinclair, C.D. and Spurr, B.D. and M.I. Ahmad , year =. Modified. Communications in Statistics --- Theory and Methods , publisher =. doi:10.1080/03610929008830405 , url =

    work page doi:10.1080/03610929008830405

  34. [34]

    Interannual variability of surface climate in the

    Marengo, Jos. Interannual variability of surface climate in the. 1992 , journal =

    1992

  35. [35]

    Further work on the prediction of northeast

    Hastenrath, Stefan and Greischar, Lawrence , year =. Further work on the prediction of northeast. Journal of Climate , volume =

  36. [36]

    1993 , booktitle =

    The relationship of drought frequency and duration to time scales , author =. 1993 , booktitle =

    1993

  37. [37]

    Grimshaw , year =

    Scott D. Grimshaw , year =. Computing maximum likelihood estimates for the generalized. Technometrics , volume =. doi:10.1080/00401706.1993.10485040 , url =

    work page doi:10.1080/00401706.1993.10485040 1993

  38. [38]

    Bias correction in

    Cordeiro, Gauss M and Klein, Ruben , year =. Bias correction in. Statistics and Probability Letters , publisher =

  39. [39]

    and Turova, Tatyana S

    Mason, David M. and Turova, Tatyana S. , year =. Weak convergence of the. Extreme Value Theory and Applications: Proceedings of the Conference on Extreme Value Theory and Applications , publisher =. doi:10.1007/978-1-4613-3638-9\_25 , isbn =

    work page doi:10.1007/978-1-4613-3638-9

  40. [40]

    1994 , journal =

    Bootstrap Recycling: A Monte Carlo Alternative to the Nested Bootstrap , author =. 1994 , journal =. doi:10.2307/2290915 , url =

    work page doi:10.2307/2290915 1994

  41. [41]

    1995 , journal =

    Controlling The False Discovery Rate---A Practical And Powerful Approach To Multiple Testing , author =. 1995 , journal =

    1995

  42. [42]

    Coles, Stuart G and Tawn, Jonathan A , year =. A. Applied Statistics , publisher =

  43. [43]

    1996 , journal =

    Second-order regular variation and rates of convergence in extreme-value theory , author =. 1996 , journal =. doi:10.1214/aop/1042644709 , url =

    work page doi:10.1214/aop/1042644709 1996

  44. [44]

    1996 , journal =

    Generalized regular variation of second order , author =. 1996 , journal =

    1996

  45. [45]

    1996 , journal =

    Excess functions and estimation of the extreme-value index , author =. 1996 , journal =

    1996

  46. [46]

    Teugels , year =

    Jan Beirlant and Petra Vynckier and Jozef L. Teugels , year =. Tail index estimation,. Journal of the American Statistical Association , publisher =. doi:10.1080/01621459.1996.10476735 , url =

    work page doi:10.1080/01621459.1996.10476735 1996

  47. [47]

    Resnick , year =

    Marie Kratz and Sidney I. Resnick , year =. The. Communications in Statistics. Stochastic Models , publisher =. doi:10.1080/15326349608807407 , url =

    work page doi:10.1080/15326349608807407

  48. [48]

    1997 , publisher =

    Bootstrap Methods and Their Application , author =. 1997 , publisher =

    1997

  49. [49]

    1997 , journal =

    Extreme deviations and applications , author =. 1997 , journal =

    1997

  50. [50]

    1997 , address =

    Regional Frequency Analysis: An Approach Based on L -Moments , author =. 1997 , address =. doi:10.1017/cbo9780511529443 , url =

    work page doi:10.1017/cbo9780511529443 1997

  51. [51]

    1997 , publisher =

    Modelling Extremal Events for Insurance and Finance , author =. 1997 , publisher =

    1997

  52. [52]

    1998 , journal =

    Comparison of tail index estimators , author =. 1998 , journal =. doi:10.1111/1467-9574.00068 , url =

    work page doi:10.1111/1467-9574.00068 1998

  53. [53]

    1998 , journal =

    Selecting the optimal sample fraction in univariate extreme value estimation , author =. 1998 , journal =. doi:10.1016/s0304-4149(98)00017-9 , url =

    work page doi:10.1016/s0304-4149(98)00017-9 1998

  54. [54]

    Estimating a tail exponent by modelling departure from a

    Andrey Feuerverger and Peter Hall , year =. Estimating a tail exponent by modelling departure from a. The Annals of Statistics , volume =. doi:10.1214/aos/1018031215 , url =

    work page doi:10.1214/aos/1018031215

  55. [55]

    1999 , journal =

    Tail index estimation and an exponential regression model , author =. 1999 , journal =. doi:10.1023/a:1009975020370 , url =

    work page doi:10.1023/a:1009975020370 1999

  56. [56]

    Methodology , author =

    Statistics for modeling heavy tailed distributions in geology: Part I. Methodology , author =. 1999 , journal =. doi:10.1023/a:1007538624271 , url =

    work page doi:10.1023/a:1007538624271 1999

  57. [57]

    1999 , journal =

    A bootstrap-based method to achieve optimality in estimating the extreme-value index , author =. 1999 , journal =. doi:10.1023/a:1009900215680 , url =

    work page doi:10.1023/a:1009900215680 1999

  58. [58]

    1999 , journal =

    Exceedances over high thresholds: a guide to threshold selection , author =. 1999 , journal =. doi:10.1023/a:1009914915709 , url =

    work page doi:10.1023/a:1009914915709 1999

  59. [59]

    Climate variability in southern

    Grimm, Alice M and Barros, Vicente R and Doyle, Moira E , year =. Climate variability in southern. Journal of Climate , volume =

  60. [60]

    2000 , journal =

    Alternatives to a Semi-Parametric Estimator of Parameters of Rare Events---The Jackknife Methodology , author =. 2000 , journal =. doi:10.1023/a:1011470010228 , url =

    work page doi:10.1023/a:1011470010228 2000

  61. [61]

    2000 , journal =

    Impacts of extreme weather and climate on terrestrial biota , author =. 2000 , journal =

    2000

  62. [62]

    Multivariate

    Wolter, Klaus , year =. Multivariate

  63. [63]

    2001 , journal =

    The Control of The False Discovery Rate in Multiple Testing Under Dependency , author =. 2001 , journal =

    2001

  64. [64]

    Coles, Stuart , year =. An. doi:10.1007/978-1-4471-3675-0 , isbn =

    work page doi:10.1007/978-1-4471-3675-0

  65. [65]

    2001 , journal =

    A predictive approach to tail probability estimation , author =. 2001 , journal =. doi:10.1023/a:1016546027962 , url =

    work page doi:10.1023/a:1016546027962 2001

  66. [66]

    2001 , journal =

    The bootstrap methodology in statistics of extremes---choice of the optimal sample fraction , author =. 2001 , journal =. doi:10.1023/a:1016592028871 , url =

    work page doi:10.1023/a:1016592028871 2001

  67. [67]

    2001 , journal =

    A diagnostic for selecting the threshold in extreme value analysis , author =. 2001 , journal =. doi:10.1111/1467-9868.00286 , url =

    work page doi:10.1111/1467-9868.00286 2001

  68. [68]

    2001 , journal =

    Using a bootstrap method to choose the sample fraction in tail index estimation , author =. 2001 , journal =. doi:10.1006/jmva.2000.1903 , url =

    work page doi:10.1006/jmva.2000.1903 2001

  69. [69]

    Goodness-of-fit tests for the generalized

    Vartan Choulakian and Michael A Stephens , year =. Goodness-of-fit tests for the generalized. Technometrics , publisher =. doi:10.1198/00401700152672573 , url =

    work page doi:10.1198/00401700152672573

  70. [70]

    2002 , journal =

    On exponential representations of log-spacings of extreme order statistics , author =. 2002 , journal =. doi:10.1023/a:1022171205129 , url =

    work page doi:10.1023/a:1022171205129 2002

  71. [71]

    2002 , journal =

    A dynamic mixture model for unsupervised tail estimation without threshold selection , author =. 2002 , journal =. doi:10.1023/a:1024072610684 , url =

    work page doi:10.1023/a:1024072610684 2002

  72. [72]

    and Best, Nicola G

    Spiegelhalter, David J. and Best, Nicola G. and Carlin, Bradley P. and Van Der Linde, Angelika , year =. Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , volume =. doi:10.1111/1467-9868.00353 , url =

    work page doi:10.1111/1467-9868.00353

  73. [73]

    2002 , journal =

    Non-parametric bootstrap recycling , author =. 2002 , journal =. doi:10.1023/a:1020754911317 , url =

    work page doi:10.1023/a:1020754911317 2002

  74. [74]

    2003 , journal =

    Inference for clusters of extreme values , author =. 2003 , journal =

    2003

  75. [75]

    2003 , journal =

    Anticipating catastrophes through extreme value modelling , author =. 2003 , journal =

    2003

  76. [76]

    Dupuis and Maria-Pia Victoria-Feser , year =

    Debbie J. Dupuis and Maria-Pia Victoria-Feser , year =. A prediction error criterion for choosing the lower quantile in

  77. [77]

    Grimm, Alice M , year =. The. Journal of Climate , volume =

  78. [78]

    Guidelines on

    Plummer, Neil and Allsopp, Terry and Lopez, Jos. Guidelines on. 2003 , publisher =

    2003

  79. [79]

    2004 , journal =

    Bayesian analysis of extreme events with threshold estimation , author =. 2004 , journal =. doi:10.1191/1471082X04st075oa , url =

    work page doi:10.1191/1471082x04st075oa 2004

  80. [80]

    2004 , journal =

    Reiss and Thomas' automatic selection of the number of extremes , author =. 2004 , journal =. doi:10.1016/j.csda.2003.11.011 , url =

    work page doi:10.1016/j.csda.2003.11.011 2004

Showing first 80 references.