ArticlePDF Available

Extreme Value Analysis : an Introduction

January 2013

Authors:

Myriam Garrido

French National Institute for Agriculture, Food, and Environment (INRAE)

Relative distance between quantiles of order p computed with N (0, 1) and Student(4) models.

…

Examples of Gumbel (γ = 0 in solid line), Fréchet (for γ = 1 in dashed line) and Reversed Weibull (for γ = −1 in dotted line) cdfs.

…

Comparing ¯ F(x) (solid line) and 1 − G 1/n

…

Excesses (Y 1 ,.. .) over a threshold u.

…

Hill estimate of the extreme value index γ against different values of k and three data sets of size n = 500 simulated from a Student distribution of parameter 3 (with a true γ = 1/3).

…

Figures - uploaded by Myriam Garrido

Content may be subject to copyright.

Content uploaded by Myriam Garrido

Content may be subject to copyright.

HAL Id: hal-00917995

https://hal-enac.archives-ouvertes.fr/hal-00917995

Submitted on 12 Dec 2013

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of sci-

entic research documents, whether they are pub-

lished or not. The documents may come from

teaching and research institutions in France or

abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est

destinée au dépôt et à la diusion de documents

scientiques de niveau recherche, publiés ou non,

émanant des établissements d’enseignement et de

recherche français ou étrangers, des laboratoires

publics ou privés.

Extreme Value Analysis : an Introduction

Myriam Charras-Garrido, Pascal Lezaud

To cite this version:

Myriam Charras-Garrido, Pascal Lezaud. Extreme Value Analysis : an Introduction. Journal de la

Societe Française de Statistique, Societe Française de Statistique et Societe Mathematique de France,

2013, 154 (2), pp 66-97. <hal-00917995>

Journal de la Société Française de Statistique

Vol. 154 No. 2 (2013)

Extreme Value Analysis: an Introduction

Titre: Introduction à l’analyse des valeurs extrêmes

Myriam Charras-Garrido1and Pascal Lezaud2

Abstract:

We provide an overview of the probability and statistical tools underlying the extreme value theory, which

aims to predict occurrence of rare events. Firstly, we explain that the asymptotic distribution of extreme values belongs,

in some sense, to the family of the generalised extreme value distributions which depend on a real parameter, called the

extreme value index. Secondly, we discuss statistical tail estimation methods based on estimators of the extreme value

index.

Résumé :

Nous donnons un aperçu des résultats probabilistes et statistiques utilisés dans la théorie des valeurs extrêmes,

dont l’objectif est de prédire l’occurrence d’événements rares. Dans la première partie de l’article, nous expliquons que

la distribution asymptotique des valeurs extrêmes appartient, dans un certain sens, à la famille des distributions des

valeurs extrêmes généralisées qui dépendent d’un paramètre réel, appelé l’indice de valeur extrême. Dans la seconde

partie, nous discutons des méthodes d’évaluation statistiques des queues basées sur l’estimation de l’indice des valeurs

extrêmes

Keywords: extreme value theory, max stable distributions, extreme value index, distribution tail estimation

Mots-clés :

théorie des valeurs extrêmes, lois max-stables, indice des valeurs extrêmes, estimation en queue de

distribution

AMS 2000 subject classiﬁcations: 60E07, 60G70, 62G32, 62E20

1. Introduction

The consideration of the major risks in our technological society has become vital because of

the economic, environmental and human impacts of industrial disasters. One of the standard

approaches to studying risks uses the extreme value theory; a branch of statistics dealing with the

extreme deviations from the median of probability distributions. Of course, this approach is based

on the language of probability theory and thus the ﬁrst question to ask is whether a probability

approach applies to the studied risk. For instance, can we use probabilities in order to study the

disappearance of dinosaurs? More recently, the Fukushima disaster, only

years after that of

Chernobyl, raises the question of the appropriateness of the probability methods used. Moreover,

as explained in Bouleau (1991), the extreme value theory aims to predict occurrence of rare events

(e.g. earthquakes of large magnitude), outside the range of available data (e.g. earthquakes of

magnitude less than

). So, its use requires some precautions, and in Bouleau (1991) the author

concludes that

The approach attributing a precise numerical value for the probability of a rare phenomenon is suspect,

unless the laws of nature governing the phenomenon are explicitly and exhaustively known [...] This does

1INRA, UR346, F-63122 Saint-Genes-Champanelle, France.

E-mail: myriam.charras-garrido@clermont.inra.fr

2ENAC, MAIAA, F-31055 Toulouse, France.

E-mail: lezaud@recherche.enac.fr

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 67

not mean that the use of probability or probability concepts should be rejected.

Nevertheless, the extreme value theory remains a well suited technique capable of predicting

extreme events. Although the application of this theory in the real world always needs to be viewed

with a critical eye, we suggest, in this article, an overview of the mathematical and statistical

theories underlying it.

As already said before, the main objective of the extreme value theory is to know or predict the

statistical probabilities of events that have never (or rarely) been observed. Firstly, the statistical

analysis of extreme values has been developed in order to study ﬂood levels. Nowadays, the

domains of application include other meteorological events (such as precipitation or wind speed),

industry (for example important malfunctions), ﬁnance (e.g. ﬁnancial crises), insurance (for very

large claims due to catastrophic events), environmental sciences (like concentration of ozone in

the air), etc.

Formally, we consider the sample

X1,...,Xn

independent and identically distributed (iid)

random variables with common cumulative distribution function (cdf)

. We deﬁne the ordered

sample by

X1,n≤X2,n≤.. . ≤Xn,n=Mn

. We are interested in two related problems. The ﬁrst

one consists in estimating the tail of the survivor function

F=1−F

: given

hn>Mn

, we want

to estimate

p=¯

F(hn)

. This corresponds to the estimation of the risk to get out a zone, for

example the probability to exceed the level of a dyke for a ﬂood application. The second problem

consists in estimating extreme quantiles: given

pn<1/n

, we want to estimate

h=¯

F−1(pn)

. This

corresponds to estimating the limit of a critical zone, as the level of a dyke for a ﬂood application,

to be exceeded with probability

. Note that since we are interested to extrapolate outside the

range of available observations, we have to assume that the quantile probability depends on

and

that limn→∞pn=0.

In both problems, the same difﬁculty arises: the cdf

is unknown and difﬁcult to estimate

beyond observed data. We want to get over the maximal observation

, that means to extrapolate

outside the range of the available observations. Both parametric and non parametric usual estima-

tion methods fail in this case. For the parametric method, models considered to give similar results

in the sample range can diverge in the tail. This is illustrated in Figure 1that presents the relative

difference between quantiles from a Gaussian and a Student distribution. For the non parametric

method,

1−b

Fn(x) = 0

x>Mn

, where

denotes the empirical distribution function, i.e. it is

estimated that outside the sample range nothing is likely to be observed. As we are interested in

extreme values, an intuitive solution is to use only extreme values of the sample that may contain

more information than the other observations on the tail behaviour. Formally, this solution leads

to a semi-parametric approach that will be detailed later.

Before starting with the description of the estimation procedures, we need to introduce the

probability background which is based on the elegant theory of max-stable distribution functions,

the counterpart of the (alpha) stable distributions, see Feller (1971). The stable distributions are

concerned with the limit behaviour of the partial sum

Sn=X1+X2+···+Xn

, as

n→∞

, whereas

the theory of sample extremes is related to the limit behaviour of

. The main result is the Fisher-

Tippett-Gnedenko Theorem 2.3 which claims that

, after proper normalisation, converges in

distribution to one of three possible distributions, the Gumbel distribution, the Fréchet distribution,

or the Reversed Weibull distribution. In fact, it is possible to combine these three distributions

together in a single family of continuous cdfs, known as the generalized extreme value (GEV)

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

68 M. Charras-Garrido and P. Lezaud

FIG URE 1. Relative distance between quantiles of order p computed with N(0,1)and Student(4) models.

distributions. A GEV is characterized by a real parameter

, the extreme value index, as a stable

distribution is it by a characteristic exponent

α∈]0,2]

. Let us mention the similarity with the

Gaussian Law, a stable distribution with

α=2

, and the Central Limit Theorem. Next we have

to ﬁnd some conditions to determine for a given cdf

the limiting distribution of

. The best

tools suited to address that are the tail quantile function (cf. (3) for the deﬁnition) and the slowly

varying functions. Finally, these results will be widened to some stationary time series.

The paper is articulated in two main Sections. In Section 2, we will set up the context in order

to state the Fisher-Tippett-Gnedenko Theorem in Subsection 2.1. In this paper, we will follow

closely the approach presented in Beirlant et al. (2004b), which transfers the convergence in

distribution to the convergence of expectations for the class of real, bounded and continuous

functions. Other recent texts include Embrechts et al. (2003) and Reiss and Thomas (1997). In

Subsection 2.2, some equivalent conditions in terms of

will be given, since it is not easy to

compute the tail quantile function. Finally, in Subsection 2.3 the condition about the independence

between the

will be relaxed in order to adapt the previous result for stationary time series

satisfying a weak dependence condition. The main result of this part is Theorem 2.12.

Section 3addresses the statistical point of view. Subsection 3.1 gives asymptotic properties of

extreme order statistics and related quantities and explains how they are used for this extrapolation

to the distribution tail. Subsection 3.2 presents tail and quantile estimations using these extrapola-

tions. In Subsection 3.3, different optimal control procedures on the quality of the estimates are

explored, including graphical procedures, tests and conﬁdence intervals.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 69

2. The Probability theory of Extreme Values

Let us consider the sample

X1,...,Xn

iid random variables with common cdf

. We deﬁne

the ordered sample by

X1,n≤X2,n≤.. . ≤Xn,n=Mn

, and we are interested in the asymptotic

distribution of the maxima Mnas n→∞. The distribution of Mnis easy to write down, since

P(Mn≤x) = P(X1≤x,...,Xn≤x) = Fn(x).

Intuitively extremes, which correspond to events with very small probability, happen near

the upper end of the support of

, hence the asymptotic behaviour of

must be related to the

right tail of the distribution near the right endpoint. We denote by

ω(F) = inf{x∈R:F(x)≥1}

the right endpoint of

and by

F(x) = 1−F(x) = P(X>x)

the survivor function of

. We

obtain that for all

x<ω(F)

P(Mn≤x) = Fn(x)→0

n→∞

, whereas for all

x≥ω(F)

P(Mn≤x) = Fn(x) = 1.

Thus

converges in probability to

ω(F)

n→∞

, and since the sequence

is increasing,

converge almost surely to

ω(F)

. Of course, this information is not very useful, so we want to

investigate the ﬂuctuations of

in the similar way the Central Limit Theorem (CLT) is derived

for the sum

Sn=∑iXi

. More precisely, we look after conditions on

which ensure that there

exists a sequence of numbers

{bn,n≥1}

and a sequence of positive numbers

{an,n≥1}

such

that for all real values x

PMn−bn

an≤x=Fn(anx+bn)→G(x)(1)

n→∞

, where

is a non-degenerate distribution (i.e. without Dirac mass). If (1) holds,

said to belong to the domain of attraction of

and we will write

F∈D(G)

. The problem is

twofold: (i) ﬁnd all possible (non-degenerate) distributions

that can appear as a limit in (1), (ii)

characterize the distributions

for which there exists sequences

(an)

and

(bn)

such that (1) holds.

Introducing the threshold

un=un(x):=anx+bn

gives the more understanding interpretation

of our problem, since

P(Mn≤un) = Fn(un) = 1−n¯

F(un)

nn

Hence, we need rather conditions on the tail

to ensure that

P(Mn≤un)

converges to a non-trivial

limit. The ﬁrst result you obtain is the following:

Proposition 2.1.

For a given

τ∈[0,∞]

and a sequence

(un)

of real numbers the two assertions

(i) n ¯

F(un)→τ, and (ii) P(Mn≤un)→e−τare equivalent.

Clearly, Poisson’s limit Theorem is the key behind this Proposition. Indeed, we assume for

simplicity that

0<τ<∞

and we let

Kn(un) = ∑n

i=1I{Xi>un}

; it is the number of excesses over the

threshold

in the sample

X1,...,Xn

. This quantity has a binomial distribution with parameters

and p=¯

F(un);

P(Kn(un) = k) = n

kpk(1−p)n−k.

The Poisson’s limit Theorem yields that

Kn(un)

converges in law to a Poisson distribution with

parameter τif and only if EKn(un)→τ; this is nothing but Proposition 2.1.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

70 M. Charras-Garrido and P. Lezaud

Now, let us assume that

X1>un

and consider the discrete time

T(un)

such that

X1+T(un)>un

and

Xi≤un

for all

1<i≤T(un)

, i.e.

T(un) = min{i≥1 : Xi+1>un}

. In order to hope for a limit

distribution, we will have to normalize T(un)by the factor n(so T(un)/n∈(0,1]); then

Pn−1T(un)>k/n=P(X2≤un,··· ,Xk+1≤un|X1>un) = F(un)n(k/n).

Let x>0, then for k=⌊nx⌋

P(n−1T(un)>x) = P(n−1T(un)>k/n) = (1−¯

F(un))n(k/n),

hence if

n¯

F(un)→τ

n→∞

, we have

P(n−1T(un)>x)→e−τx

, that means the excess times

are asymptotically distributed according to an exponential law with parameter

. The precise

approach of this result requires the introduction of the point process of exceedances

(Nn)

deﬁned

by:

Nn(B) =

∑

i=1

δi/n(B)I{Xi>un}=♯{i/n∈B:Xi>un},

where

is a Borel set on

(0,1]

and

δi/n(B) = 1

i/n∈B

and

else. Then we have the following

result (see Resnick (1987)):

Proposition 2.2.

Let

(un)n∈N

be threshold values tending to

ω(F)

n→∞

. Then, we have

limn→∞n¯

F(un) = τ∈(0,∞)

, if and only if

(Nn)

converges in distribution to a Poisson process

with parameter τas n →∞.

2.1. The possible limits

Hereafter, we work under the assumption that the underlying cdf

is continuous and strictly

increasing. What are the possible non-degenerate limit laws for the maxima

? Firstly, the limit

law of a sequence of random variables is uniquely determined up to changes of location and scale

(see Resnick (1987)), that means if there exists sequences (an)and (bn)such that

PXn−bn

an≤x→G(x),

then the relation

PXn−βn

αn≤x→H(x),

holds for the sequences (βn)and (αn)if and only if

lim

n→∞an/αn=σ∈[0,∞),lim

n→∞(bn−βn)/αn=µ∈R.

In that case,

H(x) = G((x−µ)/σ)

and we say that

and

are of the same type. Thus, a cdf

cannot be in the domain of attraction of more than one type of cdf.

Furthermore, the question turns out to be closely related to the following property, identiﬁed

by Fisher and Tippett (1928). Assume that the properly normalized and centred maxima

converges in distribution to Gand let n=mr, with m,n,r∈N. Hence, as n→∞, we have

Fn(amx+bm) = [Fm(amx+bm)]r→Gr(x).

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 71

From the previous discussion, it follows that there exist

ar>0

and

such that

Gr(x) = G(arx+

br); we say that the cdf Gis max-stable.

To emphasize the role played by the tail function, we deﬁne an equivalence relation between

cdfs in this way. Two cdfs

and

are called tail-equivalent if they have the same right end-point,

i.e. if ω(F) = ω(H) = x0, and

lim

x↑x0

1−F(x)

1−H(x)=A,

for some constant

. Using the previous discussion, it can be shown (see Resnick (1987)) that

F∈D(G)if and only if H∈D(G); moreover, we can take the same norming constants.

The main result of this Section is the Theorem of Fisher, Tippet and Gnedenko which charac-

terizes the max-stable distribution functions.

Theorem 2.3

(Fisher-Tippett-Gnedenko Theorem)

Let

(Xn)

be a sequence of iid random vari-

ables. If there exist norming constants

an>0

bn∈R

and some non degenerate cdf

such that

a−1

n(Mn−bn)

converges in distribution to

, then

belongs to the type of one of the following

three cdfs:

Gumbel: G0(x) = exp(−e−x),x∈R,

Fréchet: G1,α(x) = exp(−x−α),x≥0,α>0,

Reversed Weibull: G2,α(x) = exp(−(−x)−α),x≤0,α<0.

Figure 2shows the convergence of

(Mn−bn)/an

to its extreme value limit in case of a uniform

distribution U[0,1].

FIG URE 2. Plot of

P((Xn,n−bn)/an≤x) = (1+ (x−1)/n)n

for

n=5

(dashed line) and

n=10

(dotted line) and its

limit exp(−(1−x)) (solid line) as n →∞for a U[0,1]distribution F.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

72 M. Charras-Garrido and P. Lezaud

The three types of cdfs given in Theorem (2.3) can be thought of as members of a single family

of cdfs. For that, let us introduce the new parameter γ=1/αand the cdf

Gγ(x) = exp(−(1+γx)−1/γ)),1+γx>0.(2)

The limiting case

γ→0

corresponds to the Gumbel distribution. The cdf

Gγ(x)

is known as the

generalized extreme value or as the extreme value cdf in the von Mises form, and the parameter

is called the extreme value index. Figure 3gives examples of Gumbel, Fréchet and Reversed

Weibull distributions.

FIG URE 3. Examples of Gumbel (

γ=0

in solid line), Fréchet (for

γ=1

in dashed line) and Reversed Weibull (for

γ=−1in dotted line) cdfs.

Now, we will present the sketch of the Theorem’s proof, following the approach of Beirlant

et al. (2004b) which transfers the convergence in distribution to the convergence of expectations

for the class of real, bounded and continuous functions (see Helly-Bray Theorem in Billingsley

(1995)).

Let us introduce the tail quantile function

U(t):=inf{x:F(x)≥1−1/t},(3)

which is non-decreasing over the interval

[1,∞)

. Then, for any real, bounded and continuous

functions f,

Efa−1

n(Mn−bn)=nZ∞

−∞

fx−bn

anFn−1(x)dF (x),

=Zn

fU(n/v)−bn

an1−v

nn−1

dv.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 73

Now observe that

(1−v/n)n−1→e−v

, as

n→∞

, while the interval of integration extends to

[0,∞)

. To obtain a limit for the left-hand term, we can make

a−1

n(U(n/v)−bn)

convergent for all

positive

. Considering the case

v=1

suggests that

bn=U(n)

is an appropriate choice. Thereby,

the natural condition to be imposed is that for some positive function aand any u>0

lim

x→∞

U(xu)−U(x)

a(x)=h(u)exists,(C)

with the limit function

not identically equal to zero. We have the following Proposition (Propo-

sition 2.2 in Section 2.1 in Beirlant et al. (2004b))

Proposition 2.4. The possible limits in (C)are given by

(hγ(u) = cuγ−1

γγ6=0

h0(u) = clog u,

where c ≥0and γis real.

The case

c=0

has to be excluded since it leads to a degenerate limit, and the case

c>0

can be

reduced to the case

c=1

by incorporating

in the function

. Hence, we replace the condition

(C)by

lim

x→∞

U(xu)−U(x)

a(x)=hγ(u)exists,(Cγ).(4)

The above result entails that under (Cγ), we ﬁnd that with bn=U(n)and an=a(n)

Efa−1

n(Mn−bn)→Z∞

fhγ(1/v)e−vdv :=Z∞

−∞

f(u)dGγ(u),

as n→∞, with Gγgiven by (2).

If we write

a(x) = xγℓ(x)

, then the limiting condition

a(xu)/a(x)→uγ

leads to

ℓ(xu)/ℓ(x)→1

This kind of condition refers to the notion of regular variation.

Deﬁnition 2.5. A positive measurable function ℓon (0,∞)which satisﬁes

lim

x→∞

ℓ(xu)

ℓ(x)=1,u>0,

is called slowly varying at ∞(we write ℓ∈R0).

A positive measurable function

(0,∞)

is regularly varying at

∞

of index

γ∈R

(we write

h∈Rγ) if

lim

x→∞

h(xu)

h(x)=xγ,u>0.

The slowly varying functions play a fundamental role in probability theory, good references are

the books of Feller (1971), Bingham et al. (1989) and Korevaar (2004). In particular, we have the

following result due to Karamata (1933): ℓ∈R0if and only if it can be represented in the form

ℓ(x) = c(x)expZx

ε(u)

udu,

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

74 M. Charras-Garrido and P. Lezaud

where

c(x)→c∈(0,∞)

and

ε(x)→0

x→∞

. Typical examples are

ℓ(x) = (logx)β

for arbitrary

and

ℓ(x) = exp{(logx)β}

, where

β<1

. Furthermore, if

h∈Rγ

with

γ>0

, then

h(x)→∞

while for γ<0, h(x)→0, as x↑∞.

Because of their intrinsic importance, we distinguish between the three cases where

γ>0

γ<0

and the intermediate case where

γ=0

. We have the following result (see Theorem 2.3 in

Section 2.6 in Beirlant et al. (2004b))

Theorem 2.6. Let (Cγ)hold

(i)

Fréchet case:

γ>0

. Here

ω(F) = ∞

, the ratio

a(x)/U(x)→γ

x→∞

and

is of the

same regular variation as the auxiliary function

: moreover,

(Cγ)

is equivalent with the

existence of a slowly varying function ℓUfor which U(x) = xγℓU(x).

(ii)

Gumbel case:

γ=0

. The ratios

a(x)/U(x)→0

and

a(x)/{ω(F)−U(x)} → 0

when

ω(F)

is ﬁnite.

(iii)

Reversed Weibull case:

γ<0

. Here

ω(F)

is ﬁnite, the ratio

a(x)/{ω(F)−U(x)}→−γ

and

{ω(F)−U(x)}

is of the same regular variation as the auxiliary function

: moreover,

(Cγ)

is equivalent with the existence of a slowly varying function

ℓU

for which

ω(F)−U(x) =

xγℓU(x).

2.2. Equivalent conditions in terms of F

Until now, only necessary and sufﬁcient conditions on

have been given in such a way that

F∈D(Gγ)

. Nevertheless, it is not always easy to compute the tail quantile function of a cdf

So, it could be preferable to express the relation between (Cγ)to the underlying distribution F.

The link between the tail of

and its tail quantile function

depends on the concept of the de

Bruyn conjugate (see Proposition 2.5 in Section 2.9.3 in Beirlant et al. (2004b)).

Proposition 2.7. If ℓ∈R0, then there exists ℓ∗∈R0, the de Bruyn conjugate of ℓ, such that

ℓ(x)ℓ∗(xℓ(x)) →1,x↑∞.

Moreover,

ℓ∗

is asymptotically unique in the sense that if also

ℓ

is slowly varying and

ℓ(x)b

ℓ(xℓ(x)) →

1, then ℓ∗∼b

ℓ. Furthermore, (ℓ∗)∗∼ℓ.

This yields the full equivalence between the statements

1−F(x) = x−1/γℓF(x),and U(x) = xγℓU(x),

where the two slowly varying functions

ℓF

and

ℓU

are linked together via the de Bruyn conjugation.

So, according to Theorem 2.6 (i) and (iii) we get that

Theorem 2.8. Referring to the notation of Theorem 2.6, we have:

(i)

Fréchet case:

γ>0

F∈D(Gγ)

if and only if there exists a slowly varying function

ℓF

for

which

F(x) = x−1/γℓF(x)

. Moreover, the two slowly varying functions

ℓU

and

ℓF

are linked

together via the de Bruyn conjugation.

(ii)

Reversed Weibull case:

γ<0

F∈D(Gγ)

if and only if there exists a slowly varying

function

ℓF

for which

Fω(F)−x−1∼x1/γℓF(x),x↑∞

. Moreover, the two slowly varying

functions ℓUand ℓFare linked together via the de Bruyn conjugation.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 75

When the cdf

has a density

, it is possible to derive sufﬁcient conditions in terms of the

hazard function

r(x) = f(x)/(1−F(x))

. These conditions, which are due to von Mises (1975),

are known as the von Mises conditions. In particular, the calculations involved on checking

the attraction condition to

are often tedious, in this respect, the von Mises criterion can be

particularly useful.

Proposition 2.9

(von Mises’ Theorem)

Sufﬁcient conditions on the density of a distribution for

it belongs to D(Gγ)are the following:

(i) Fréchet case: γ>0. If ω(F) = ∞and limx↑∞xr(x) = 1/γ, then F ∈D(Gγ),

(ii)

Gumbel case:

γ=0

r(x)

is ultimately positive in the neighbourhood of

ω(F)

, is differen-

tiable there and satisﬁes limx↑ω(F)dr(x)

dx =0, then F ∈D(G0)

(iii)

Reversed Weibull case:

γ<0

ω(F)<∞

and

limx↑ω(F)(ω(F)−x)r(x) = 1/γ

, then

F∈

D(Gγ).

Some examples of distributions which belong to the Fréchet, the Reversed Weibull and the

Gumbel domain are given in respectively Table 1, Table 2and Table 3. For more details about the

norming constants

and

, see Embrechts et al. (2003). We also recall that the choice of these

constants is not unique, for example we can choose

αn

instead of

limn→∞an/αn=1

(see the

beginning of the Section 2.1).

TABL E 1. A list of distributions in the Fréchet domain

Distribution 1 −F(x)Extreme value

index

Pareto ∼Kx−α,K,α>01

F(m,n)R∞

Γ(m+n

Γ(m

2)Γ(n

2)ωm

2−11+m

nω−m+n

2dω

x>0; m,n>0

Fréchet 1−exp(−x−α)

x>0;α>0

TnR∞

2Γ(n+1

√nπΓ(n

2)1+ω2

n−n+1

2dω

x>0;m,n>0

TABL E 2. A list of distributions in the Reversed Weibull domain

Distribution 1 −Fω(F)−1

xExtreme value

index

Uniform

x>1−1

Beta(p,q)R1

1−1

Γ(p+q)

Γ(p)Γ(q)up−1(1−u)q−1du

x>1; p,q>0−1

Reversed Weibull 1−exp(−x−α)

x>0; α>0−1

Finally, we give an alternative condition for

(Cγ)

(Proposition 2.1 in Section 2.6 in Beirlant

et al. (2004b)). It constitutes the basis for numerous statistical techniques to be discussed in

Section 3.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

76 M. Charras-Garrido and P. Lezaud

TABL E 3. A list of distributions in the Gumbel domain

Distribution 1 −F(x)

Weibull exp(−λxτ),x>0; λ,τ>0

Exponential exp(−λx),x>0; λ>0

Gamma λm

Γ(m)R∞

xum−1exp(−λu)du,x>0; α,m>0

Logistic 1/(1+exp(x)),x∈R

Normal R∞

√2πσ 2exp −(x−µ)2

2σ2,x∈R;σ>0,µ∈R

Log-normal R∞

√2πσ 2uexp −1

2σ2(logu−µ)2du,x>0; µ∈R,σ>0

Proposition 2.10.

The distribution

belongs to

D(Gγ)

if and only if for some auxiliary function

b and 1+γv>0

1−F(y+b(y)v)

1−F(y)→(1+γv)−1/γ,(C∗

γ)

as y →ω(F). Then

b(y+vb(y))

b(y)→1+γv.

Condition (C∗

γ)has an interesting probabilistic interpretation. Indeed, (C∗

γ)reformulates as

lim

x↑ω(F)

PX−v

b(v)>xX>v= (1+γv)−1/γ.

Hence, the condition

(C∗

γ)

gives a distributional approximation for the scaled excesses over

the high threshold

, and the appropriate scaling factor is

b(v)

. This motivates the following

deﬁnitions.

Let Xbe a random variable with cdf Fand right endpoint ω(F). For a ﬁxed u<ω(F),

Fu(x) = P(X−u≤x|X>u),x≥0 (5)

is the excess cdf of the random variable Xover the threshold u. The function

e(u) = E(X−u|X>u)

is called the mean excess function of

. The function

uniquely determines

. Indeed, whenever

Fis continuous, we have

1−F(x) = e(0)

e(x)exp−Zx

e(u)du,x>0.

Deﬁne the cdf Hγby

Hγ(x) = (1−(1+γx)−1/γ,if γ6=0,

1−e−x,if γ=0,

where

x≥0

γ≥0

and

0≤x≤ −1/γ

γ<0

Hγ

is called a standard generalised Pareto

distribution (GPD). In order to take into account a scale factor σ, we will denote

Hγ,σ(x) = (1−(1+γ(x/σ))−1/γ,if γ6=0,

1−e−x/σ,if γ=0,(6)

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 77

which is deﬁned for

x∈IR+

γ≥0

and

x∈[0,−σ/γ[

γ<0

. Then, condition

(C∗

γ)

above

suggests a GPD as appropriate approximation of the excess cdf

for large

. This result is often

formulated as follows in Pickands (1975): for some function σto be estimated from the data

Fu(x)≈Hγ,σ(u)(x).

2.3. Extremes of Stationary Time Series

Beforehand, we restricted ourselves to iid random variables. However, in reality extremal events

often tend to occur in clusters caused by local dependence. This requires a modiﬁcation of standard

methods for analysing extremes. We say that the sequence of random variable

(Xi)

is strictly

stationary if for any integer

h≥0

and

n≥1

, the distribution of the random vector

(Xh+1,...,Xh+n)

does not depend on

. We seek the limiting distribution of

(Mn−bn)/an

for some choice of

normalizing constants

an>0

and

. However, the limit distribution needs not to be the same as

for the maximum

of the associated independent sequence

Xi)1≤i≤n

with the same marginal

distribution as

(Xi)

. For instance, starting with an iid sequence

(Yi,1≤i≤n+1)

of random

variables with common cdf

, we deﬁne a new sequence of random variables

(Xi,1≤i≤n)

Xi=max(Yi,Yi+1). We see that the dependence causes large values to occur in pairs. Indeed, the

random variables

are distributed according to the cdf

F=H2

; so if

satisﬁes the equivalent

conditions in Proposition 2.1, we conclude that

nH(un)→τ/2

. Consequently, the maximum

Mn=Xn,nsatisﬁes

lim

n→∞

P(Mn≤un) = e−τ/2.

To hope for the existence of a limiting distribution of

(Mn−bn)/an

, the long-range dependence

at extreme levels needs to be suitably restricted. To measure the long-range dependence, Leadbetter

(1974) introduced a weak dependence condition known as the

D(un)

condition. Before setting out

this condition, let us introduce some notations as in Beirlant et al. (2004b). For a set

of positive

integers, let

M(J) = maxi∈JXi

(with

M(/0) = −∞)

). If

I={i1,...,ip}

J={j1,..., jq}

, we write

that I≺Jif and only if

1≤i1<··· <ip<j1·· · <jq≤n,

and the distance d(I,J)between Iand Jis given by d(I,J) = j1−ip.

Condition 2.11

(

D(un)

)

For any two disjoint subsets

{1,...,n}

such that

I≺J

and

d(I,J)≥lnwe have

P({M(I)≤un}∩{M(J)≤un})−P(M(I)≤un)P(M(J)≤un)≤αn,ln

and αn,ln→0 as n→∞for some positive integer sequence lnsuch that ln=o(n).

The

D(un)

condition says that any two events of the form

{M(I)≤un}

and

{M(J)≤un}

become asymptotically independent as

increases when the index sets

and

are separated by

a relatively short distance

ln=o(n)

. This condition is much weaker than the standard forms of

mixing condition (such as strong mixing).

Now, we partition the integers

{1,...,n}

into

disjoint blocks

Ij={(j−1)rn+1,..., jrn}

size

rn=o(n)

with

kn= [n/rn]

and, in case

knrn<n

, a remainder block,

Ikn+1={knrn+1,...,n}

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

78 M. Charras-Garrido and P. Lezaud

A crucial point is that the events

{Xi>un}

are sufﬁciently rare for the probability of an exceedance

occurring near the ends of the blocks

to be negligible. Therefore, if we drop out the remainder

block and the terminal sub-blocks

I′

j={jrn−ln+1,..., jrn}

of size

, we can consider only the

sub-block

I∗

j={(j−1)rn+1,..., jrn−ln}

which are approximatively independent. Thus we get

P(Mn≤un) = P kn

j=1{M(I∗

j)≤un}!+o(1).

Finally, using condition D(un)with knαn,ln→0, we obtain

P kn

j=1{M(I∗

j)≤un}!−Pkn({M(I∗

1)≤un})≤knαn,ln→0,

n→∞

. Now, we observe that if thresholds

increase at a rate such that

limsup nF (un)>∞

then

Pkn(M(I∗

1)≤un)−Pkn(Mrn≤un)≤kn|P(M(I∗

1)≤un)−P(Mrn≤un)|

=knP(M(I∗

1)≤un<M(I′

1))

≤knlnP(X1>un)→0.

So, under the D(un), we obtain the appropriate condition

P(Mn≤un)−Pkn(Mrn≤un)→0 (7)

from which the following fundamental results were derived, see Leadbetter (1974,1983)

Theorem 2.12.

Let

(Xn)

be a stationary sequence for which there exist sequences of constants

an>0and bnand a non-degenerate distribution function G such that

PMn−bn

an≤x→G(x),n→∞.

D(un)

holds with

un=anx+bn

for each

such that

G(x)>0

, then

is an extreme value

distribution function.

Theorem 2.13.

If there exist sequences of constants

an>0

and

and a non-degenerate distri-

bution function e

G such that

P e

Mn−bn

an≤x!→e

G(x),n→∞,

D(anx+bn)

holds for each

such that

G(x)>0

and if

P[(Mn−bn)/an≤x]

converges for some

x, then we have

PMn−bn

an≤x→G(x):=e

Gθ(x),n→∞,

for some constant θ∈[0,1].

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 79

Theorem 2.12 shows that the possible limiting distributions for maxima of stationary sequences

satisfying the

D(un)

condition are the same as those for maxima of independent sequences.

Nevertheless Theorem 2.12 does not mean that the relations

Mn∈D(G)

and

Mn∈D(e

hold

with

G=G

. In fact,

is often of the form

Gθ

for some

θ∈[0,1]

(see for instance the introductory

example). This is precisely what Theorem 2.13 claims.

The constant

is called extremal index and always belongs to the interval

[0,1]

. For instance,

if we consider the max-autoregressive process of order one deﬁned by the recursion

Xi=max{αXi−1,(1−α)Zi}

where

0≤α<1

and where the

are independent Fréchet random variables. Then it can be

proved that (cf. Beirlant et al. (2004b) Section 10.2.1)

P(Mn≤x) = P(X1≤x)[P(Z1≤x/(1−α)] →exp[−(1−α)/x]:=G(x)

Whereas,

G(x) = exp(−1/x)

, so

θ=1−α

. This example shows that any number in

(0,1]

can

be an extremal index. The case

θ=0

is pathological, it entails that sample maxima

of the

process are of smaller order than sample maxima

. We refer to Leadbetter et al. (1983) and

Denzel and O’Brien (1975) for some examples. Moreover,

θ>1

is impossible; this follows from

the following argument (see Embrechts et al. (2003) Section 8.1.1):

P(Mn≤un) = 1−P n

[

i=1{Xi>un}!≥1−nF(un).

The left-hand side converges to

e−θτ

whereas the right-hand side has limit

1−τ

, hence

e−θτ ≥

1−τ

for all

τ>0

which is possible only if

θ≤1

. A case in which there is no extremal index is

given in O’Brien (1974). In this article, each

is uniform over

[0,1]

X1,X3,. ..

being independent

and

X2n

a certain function of

X2n−1

for each

. Finally, a case where

D(un)

does not hold but the

extremal index exists is given by the following example of Davis (1982). Let

Y1,Y2, . ..

, be iid, and

deﬁne the sequence

(X1,X2,X3, .. .) = (Y1,Y2,Y2,Y3,Y3,. . .)or (Y1,Y1,Y2,Y2,. ..)

each with probability

1/2

. It follows from Davis (1982) that the sequence

(Xn)

has extremal

index

1/2

. However

D(un)

does not hold: for example, if

X1=X2

then

Xn=Xn+1

is odd and

Xn6=Xn+1if nis even. For more details, we refer to Leadbetter (1983).

To sum up, unless

is equal to one, the limiting distributions for the independent and stationary

sequences are not the same. Moreover, if

θ>0

then

is an extreme value distribution, but with

different parameters than e

G. Thus if

G(x) = exp −1+γx−µ

σ−1/γ!,

then we have

G(x) = exp −1+γx−e

σ−1/γ!,

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

80 M. Charras-Garrido and P. Lezaud

with µ=e

µ−e

σ(1−θγ)/γand σ=e

σθ γ(if γ=0, σ=e

σand µ=e

µ+σlogθ).

Under some regularity assumptions the limiting expected number of exceedances over

a block containing at least one such exceedance is equal to

1/θ

(if

θ>0

). In fact, using the

notations previously introduced, we obtain (see Beirlant et al. (2004b) Section 10.2.3)

θ=lim

n→∞

rn¯

F(un)

P(Mrn>un)=E"rn

∑

i=1

1(Xi>un)Mrn>un#.

We can have an insight into this result with the following approach: let us assume that

is a

threshold sequence such that

nF(un)→τ

and

P(Mn≤un)→exp(−θτ)

, then from (7) (with

kn=⌊n/rn⌋) we get n

P(Mrn>un)→θτ,

and conclude that

θ=lim

n→∞

P(Mrn>un)

rnF(un).

Another interpretation of extremal event, due to O’Brien (1987), is that under some assumptions

represents the limiting probability that an exceedance is followed by a run of observations

below the threshold

θ=lim

n→∞

P(max{X2,X3,...,Xrn} ≤ un|X1>un).

So, both interpretations identify

θ=1

with exceedances occurring singly in the limit, unlike

θ<1 which implies that exceedances tend to occur in clusters.

The case

θ=1

can be checked by using the following sufﬁcient condition

D′(un)

introduced

by Leadbetter (1974), when allied with D(un),

Condition 2.14 (D′(un)).

lim

k→∞limsup

n→∞

n⌊n/k⌋

∑

j=2

P(X1>un,Xj>un) = 0.

Notice that D′(un)implies

E"∑

1≤i<j≤⌊n/k⌋

1(Xi>un,Xj>un)#≤ ⌊n/k⌋⌊n/k⌋

∑

j=2

E[1(X1>un,Xj>un)→0],

so that, in the mean, joint exceedances of unby pairs (Xi,Xj)become unlikely for large n.

Verifying the conditions

D(un)

and

D′(un)

is, in general, tedious, except in the case of a

Gaussian stationary sequence. Indeed, let

r(n) = cov(X0,Xn)

be the auto-covariance function,

then the so called Berman’s condition r(n)logn→∞allied with lim supn→∞nΦ(un)<∞, where

is the normal distribution, are sufﬁcient to imply the both conditions

D(un)

and

D′(un)

(see

Leadbetter et al. (1983)). Let recall that the normal distribution

is in the Gumbel maximum

domain of attraction.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 81

3. The Statistical point of view of Extreme Values Theory

As mentioned in the Introduction, the cdf

is unknown and difﬁcult to estimate beyond observed

data, so we need to extrapolate outside the range of the available observations. In this Section,

using the properties developed in Section 2, we will introduce and discuss different procedures

capable of carrying out this extrapolation.

3.1. Extrapolation to the distribution tail

Firstly, we can use the properties of the maximum

given in Section 2for this extrapolation, as

presented in Subsection 3.1.1. We can also base our extrapolation to the distribution tail on the

excesses or peaks over a threshold as presented in Subsection 3.1.2. Both extrapolation procedures

are derived from asymptotic procedures that correspond to a ﬁrst order approximation of the

distribution tail. Second order conditions as presented in Subsection 3.1.3 may help to improve

this approximation.

3.1.1. Using maxima

Theorem 2.3 gives the asymptotic distribution of the maximum

. Then we use the approximation

of the distribution of Mnby the generalized extreme value (GEV) cdf (2) to write

F(x) = P(Mn≤x)1/n∼G1/n

γx−bn

an,x→ω(F).

This gives a semi-parametric approximation of the tail of the cdf

. This approximation is

illustrated in Figure 4for a uniform distribution and different values of

, using the theoretical

values of

and

. Let recall that the uniform distribution is in the Reversed Weibull maximum

domain of attraction, and that in this case γ=−1 (cf. Table 2).

We can equivalently approximate an extreme quantile by

F−1(pn) = U(1/pn)∼bn+an

γ(−ln(1−pn)n)−γ−1,pn→0 when n→∞.

In these two approximations, appear three quantities

and

whose theoretical values are

only known when the cdf

is known. In practice, these quantities are unknown.

corresponds

to a shape parameter,

to a scale parameter, and

is the extreme value index. These parameters

would be estimated in Subsection 3.2 to produce semi-parametric estimations of the distribution

tail. In this case, this estimation would be performed using a block maxima sample.

3.1.2. Using Peaks Over a Threshold: The POT method

Modelling block maxima is a wasteful approach to extreme value analysis if other data on extremes

are available. A natural alternative is to regard observations that exceed some high threshold

smaller than the right endpoint ω(F)of Fas extreme events.

Excesses occur conditioned on the event that an observation is larger than a threshold

. They

are denoted by

(Y1, . ..)

and represented in Figure 5. The excess cdf

deﬁned in (5) expresses

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

82 M. Charras-Garrido and P. Lezaud

FIG URE 4. Comparing

F(x)

(solid line) and

1−G1/n

−1((x−bn)/an)

with

an=n−1

and

bn=1

for

n=50

(dashed line)

and n =100 (dotted line) for a uniform distribution F.

also as

Fu(y) = P(X≤u+y|X≥u) = 1−F(u+y)

F(u),y>0.

Pickands’ Theorem (Pickands (1975)) implies that

can be approximated by a generalized

Pareto distribution (GPD) function given by (6). Parameter

is the extreme value index, and

σ=an+γ(u−bn)

. In Section 3.1.1, approximating the distribution of the maximum by an

EVD leads to semi-parametric estimations of the tail of the cdf

and an extreme quantile.

Equivalently, approximating the distribution of the excesses over a threshold

may lead to the

following semi-parametric approximations. For the tail of the cdf

, we have the semi-parametric

approximation

F(x)∼1−F(u)Hγ,σ(x−u)

x→ω(F)

. And for an extreme quantile, we obtain

the semi-parametric approximation

F−1(pn)∼u+σ

γ"p

F(u)−γ

−1#,pn→0 when n→∞.

Again, we have three unknown parameters

and

to be estimated (see Subsection 3.2).

Note that in practice,

u<Mn

corresponds to a quantile inside the sample range that can be

easily estimated by an observation (a quantile of the empirical distribution function). In practice,

we choose

u=Xn−k+1,n

, where

is the number of excesses. However, this does not avoid

the estimation of a parameter since

has to be accurately chosen. This choice is detailed in

Subsection 3.3.1.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 83

FIG URE 5. Excesses (Y1, . . .)over a threshold u.

3.1.3. Second order conditions

The ﬁrst order condition (

Cγ

), or equivalently (

C∗

), relies to the convergence in distribution of

the maximum

. We are now interested in the convergence rate for the distribution of the

maximum

to the extreme value distribution. It corresponds to derive a remainder (see for

example de Haan and Ferreira (2006) Section 2.3 or Beirlant et al. (2004a) Section 3.3) of the

limit expressed by the ﬁrst order condition (Cγ).

The function

(or the corresponding probability distribution) is said to satisfy the second

order condition if for some positive function

and some positive or negative function

with

limt→∞A(t) = 0

lim

t→∞

U(tx)−U(t)

a(t)−xγ−1

A(t)=Ψ(x),x>0,(8)

where

is some function that is not a multiple of the function

(xγ−1)/γ

. Functions

and

are

sometimes referred to as respectively ﬁrst order and second order auxiliary functions. However,

note that for

identically one, we obtain the ﬁrst order condition (

Cγ

) with

identically zero. The

second order condition has been used to prove the asymptotic normality of different estimators

and to deﬁne some of the estimators detailed in the following Section.

The following result (see de Haan and Ferreira (2006) Section 2.3) gives more insights on the

functions a,Aand Ψ.

Theorem 3.1.

Suppose that the second order condition (8) holds. Then there exists constants

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

84 M. Charras-Garrido and P. Lezaud

c2∈IR and some parameter ρ≤0such that

Ψ(x) = c1Zx

sγ−1Zs

uρ−1dud s +c2Zx

sγ+ρ−1.(9)

Moreover, for x >0,

lim

t→∞

a(tx)

a(t)−xγ

A(t)=c1xγxγ−1

γ(10)

and

lim

t→∞

A(tx)

A(t)=xρ.(11)

Equation (11) means that function

is regularly varying with index

, while equation (10)

gives a link between functions aand A. For ρ6=0, the limiting function Ψcan be expressed as

Ψ(X) = c1

ρxγ+ρ−1

γ+ρ−xγ−1

γ+c2

xγ+ρ−1

γ+ρ.

If ρ=0 and γ6=0, Ψcan be written as

Ψ(X) = c1

γxγlog(x)−xγ−1

γ+c2

xγ−1

γ.

Finally, for ρ=0 and γ=0, Ψcan be written as

Ψ(X) = c1

2(log(x))2+c2log(x).

There are several equivalent expressions for these quantities that can be found e.g. in de Haan and

Ferreira (2006) Section 2.3 or Beirlant et al. (2004a) Section 3.3.

3.2. Estimation

We present the estimation procedure both for the block maxima and peak over threshold methods.

Thus, in order to be general, we express the estimates from the original sample (

X1,...,Xn

). We

detail different estimates including maximum likelihood, moment, Pickands, Hill, regression

and Bayesian estimates. In all cases, we focus on estimating the extreme value index

. Other

parameters can be deduced and are not detailed.

3.2.1. Maximum likelihood estimates

Maximum likelihood is usually one of the most natural estimates, largely used owing to its good

properties and simple computation. However, in the case of extreme estimates, the support of the

EVD (or the GPD) depends on the unknown parameter values. Then, as detailed by Smith (1985),

the usual regularity conditions underlying the asymptotic properties of maximum likelihood

estimators are not satisﬁed. In case

γ>−1/2

, the usual properties of consistency, asymptotic

efﬁciency and asymptotic normality hold. But there is no analytic expression for the maximum

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 85

likelihood estimates. Then, maximization of the log-likelihood may be performed by standard

numerical optimization algorithms, see e.g. Prescott and Walden (1980,1983), Hosking (2013) or

Macleod (1989). An iterative formula is also available and presented in Castillo et al. (2004).

Moreover, remark that standard convergence properties are valuable for estimating using a

sample issued from an EVD (or a GPD). Nevertheless, Fisher-Tippet-Gnedenko Theorem 2.3

(or Pickands’ Theorem in Pickands (1975)) only guarantees that the maximum

(or the peaks

over threshold) is approximately EVD (or GPD). Their accuracy in the context of extremes is

more difﬁcult to assess. However, asymptotic normality has been ﬁrst proved for

γ>−1/2

see e.g. de Haan and Ferreira (2006) Section 3.4. More recently, Zhou (2009,2010) proves the

asymptotic normality for

γ>−1

and the non-consistency for

γ<−1

. Conﬁdence intervals follow

immediately from this approximate normality of the estimator. But these properties are limited to

the range

γ>−1

concerning the quantity to be estimated. In practice, the potential range of value

of the parameter is unknown and thus the accuracy of the estimation cannot be assessed. Then,

alternative estimates have been proposed.

3.2.2. Moment and probability weighted moment estimates

The probability weighted moments of a random variable

with cdf

, introduced by Greenwood

et al. (1979), are the quantities

Mp,r,s=E(XpFr(X)(1−F(X))s)

, for real

and

. The standard

moments are obtained for

r=s=0

. Moments and probability weighted moments do not exist for

γ≥1. For γ<1, we obtain for the EVD, setting p=1 and s=0,

M1,r,0=1

r+1b−a

γ[1−(r+1)γΓ(1−γ)],

and for the GPD, setting p=1 and r=0,

M1,0,s=σ

(s+1)(s+1−γ).

By estimating these moments from a sample of block maxima or excesses over a threshold, we

obtain estimates of the parameters

a,b,σ,γ

. Note that for block maxima and EVD, there is no

analytic expression for the estimate of

that has to be computed numerically. Conversely, for

peaks over threshold and GPD, we have the following analytic expression given in Hosking and

Wallis (1987)

γPWM (k) = 2−b

M1,0,0

M1,0,0−2b

M1,0,1

with b

M1,0,s=1

∑

i=11−i

k+1s

Yi,n.

Its conceptual simplicity, its easy implementation and its good performance for small samples

make this approach still very popular. However, this does not apply for strong heavy tails and in

this case again, the range limitation

γ<1

concerns the quantity to be estimated. Moreover, the

asymptotic normality is only valid for

γ∈]−1,1/2[

, see Hosking and Wallis (1987) or de Haan

and Ferreira (2006) Section 3.6.

To overcome these drawbacks, generalized probability weighted moment estimates have been

proposed by Diebolt et al. (2007) for the parameters of the GDP distribution that exist for

γ<2

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

86 M. Charras-Garrido and P. Lezaud

and are asymptotically normal for

γ∈]−1,3/2[

.Diebolt et al. (2008) also proposed generalized

probability weighted moment estimates for the parameters of the EVD distribution that exist for

γ<b+1

and are asymptotically normal for

γ<1/2+b

, for some

b>0

. However, since these

are usual estimates, as the maximum likelihood estimate, they were not been speciﬁcally designed

for extreme modelling. Conversely, the following estimates have been proposed in the context of

extreme values.

3.2.3. Hill and moment estimates

Let γ>0, i.e. we place in the Fréchet domain of attraction. From (i) in Theorem 2.8, we have

lim

t→∞F(tx)/F(t) = x−1/γ,for x>1.

This means that the distribution of the relative excesses

Xi/t

over a high threshold

conditionally

Xi>t

is approximately Pareto:

P(X/t>x|X>t)≃x−1/γ

for

large and

x>1

. The likelihood

equation for this Pareto distribution leads to the Hill’s estimator (Hill (1975))

γH(k) = 1

∑

i=1

(logXn−i+1,n−logXn−k,n).

We can also remark that, for

γ>0

, an exponential quantile plot based on log-transformed data

(also called generalized quantile plot) is ultimately linear with slope

near the largest observations.

This regression point of view also leads to the Hill estimate. This estimator can also be expressed

as a simple average of scaled log-spacings

Zi=i(logXn−i+1,n−logXn−i,n),j=1,...,k.(12)

The Hill estimate is designed from the extreme value theory and is consistent, see Mason

(1982). It is also asymptotically normally distributed with mean

and variance

γ2/k

, see e.g.

Beirlant et al. (2004a) Sections 4.2 and 4.3. Conﬁdence intervals immediately follow from this

approximate normality. But the deﬁnition of the Hill estimates and its properties are again limited

to some ranges of

,i.e.

γ>0

. Moreover, in many instances a severe bias can appear related

to the slowly varying part in the Pareto approximation. Furthermore, as many estimators based

on log-transformed data, the Hill estimator is not invariant to shifts of the data. And as for all

estimates of

, for every choice of

, we obtain a different estimator, that can be very different in

the case of the Hill estimator (see Figure 6).

The moment estimator has been introduced by Dekkers et al. (1989) as a direct generalization

of the Hill estimator:

γM(k) = ˆ

γH(k) + 1−1

2 1−ˆ

γH(k)

H(2)

k!,

with

H(2)

k=1

∑

i=1

(logXn−i+1,n−logXn−k,n)2.

This estimate is deﬁned for

γ∈IR

and is consistent. But it converges in probability to

only

for

γ≥0

, see Beirlant et al. (2004a) Section 5.2. Under appropriate conditions including the

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 87

second-order condition, the asymptotic normality is established in Dekkers et al. (1989) and

recalled for example in de Haan and Ferreira (2006) Section 3.5. It can be noted that the moment

estimator is a biased estimator of γ.

3.2.4. Other regression estimates

The problem of non smoothness of the Hill estimate as a function of

can be solved with the

partial least-squares regression procedure that minimizes with respect to δand γ

∑

i=1logXn−i+1,n−δ+γlog n+1

i2

This leads to the Zipf estimate, see e.g. Beirlant et al. (2004a) Section 4.3:

γ+

Z(k) =

k∑k

i=1log k+1

i−1

k∑k

j=1k+1

jlogXn−i+1,n

k∑k

i=1log2k+1

i−1

k∑k

i=1log k+1

i2.

The asymptotic properties of this estimator are given e.g. in Csorgo and Viharos (1998).

Other reﬁnements make use of the Hill estimate through

UHi,n=Xn−i,nˆ

γH(i)

to reduce bias

and to increase smoothness as a function of

. Using these UH statistics instead of the ordered

statistics, the slope in the generalized quantile plot is estimated by (see Beirlant et al. (2004a)

Section 5.2)

γH(k) = 1

∑

i=1

(logU Hi,n−logUHk+1,n),

and another Zipf estimate based on unconstrained least square regression (see Beirlant et al. (2002,

2004a), Section 5.2)

γZ(k) =

k∑k

i=1log k+1

i+1−1

k∑k

j=1k+1

j+1logU Hi,n

k∑k

i=1log2k+1

i+1−1

k∑k

i=1log k+1

i+12.

One of the main interests of this last estimator is its smoothness as a function of

, which in some

sense reduces the difﬁcult problem of choosing k(detailed in Section 3.3.1).

Concerning the shift problems of the Hill estimate, a location-invariant variant is proposed in

Fraga Alves (2002) using a secondary k-value denoted by k0(<k)

γ(H)(k0,k) = 1

∑

i=1

log Xn−i+1,n−Xn−k,n

Xn−k0,n−Xn−k,n

This estimator is consistent and asymptotically normal with mean

and variance

γ2/k0

. Thus, its

variance is not increased drastically compared to the Hill estimator.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

88 M. Charras-Garrido and P. Lezaud

3.2.5. Pickands Estimator

Condition (Cγ) given in equation (4) leads to

log2 log U(4y)−U(2y)

U(2y)−U(y)∼γ,for large y.

Taking

y= (n+1)/k

and replacing

U(x)

by its empirical version

Un(x) = Xn−⌈n/x⌉+1,n

yields the

Pickands estimator in Pickands (1975):

γP(k) = 1

log2 log Xn−⌈k/4⌉+1,n−Xn−⌈k/2⌉+1,n

Xn−⌈k/2⌉+1,n−Xn−k+1,n.

The Pickands estimator is very simple but has a rather large asymptotic variance, see Dekkers and

de Haan (1989). Moreover, as the Hill estimate, its is amply varying as a function of

. This is

a problem as it makes crucial the choice of the fraction sample

to use for extreme estimation.

Different variants have been proposed, see e.g. Segers (2005).

3.2.6. Bayesian estimates

An alternative to frequentist estimation, as presented until now, is to proceed to a Bayesian

estimation. Some Bayesian estimates have been proposed in the literature and a review can be

ﬁnd e.g. in Coles and Powell (1996) or Coles (2001), Section 9.1. These estimators are also

still under study: more recent articles present new Bayesian estimates for extreme values. For

example, Stephenson and Tawn (2004) propose to estimate the parameters of the GPD distribution

given the domain of attraction i.e. with constraints on parameter γ.Diebolt et al. (2005) propose

quasi-conjugate Bayesian estimates for the parameters of the GPD distribution in the context of

heavy tails i.e. for

γ>0

.do Nascimento et al. (2011) are concerned with extreme value density

estimation using POT method and GPD distributions.

In our context of extreme values analysis, data are often scarce since we have to take into

account only extreme data, i.e. a small fraction

of the original sample. One of the main reasons

to use Bayesian estimation is the facility to include other sources of information through the

chosen prior distribution. This can be particularly important in the context of extremes given

the lack of information and the uncertainty in extrapolation. Moreover, the output of a Bayesian

analysis, the posterior distribution, directly gives a measure of parameter uncertainty that allows

to quantify the uncertainty in prediction. However, a Bayesian estimation implies the choice of

a prior distribution that can greatly inﬂuence the result. Thus, this adds another choice to the

determination of an adequate sample fraction k(detailed in Section 3.3.1).

3.2.7. Reducing bias

Classical extreme value index estimators are known to be quite sensitive to the number

top order statistics used in the estimation. The recently developed second order reduced-bias

estimators show much less sensitivity to changes in

, making the choice of

less crutial and

allowing to use more data for extreme estimation. These estimators are based on the second order

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 89

condition presented in Section 3.1.3. Many of them use an exponential representation including

second order parameters.

Beirlant et al. (2004a)[Section 4.4], details that, for

γ>0

the scaled log-spacings

, deﬁned

in equation (12), are approximately exponentially distributed with mean

γ+ (k/j)ρbn,k

. This

implies that estimating

from the log-spacing

, as done with the Hill estimate, leads to a bias

that is controlled by

bn,k

. In the general case, it can be shown, as presented in Beirlant et al.

(2004a)[Section 5.4], that the log-ratio spacings

jlog Xn−j+1,n−Xn−k,n

Xn−j,n−Xn−k,n

,j=1,...,k−1

are approximately exponentially distributed with mean

γ/(1−(j−(k+1))γ)

. A joint estimate

bn,k

and

computed from these properties, or variations of it, produces estimates of

with

reduced bias for heavy tail distributions or in the general case. Different proposals are presented

in Beirlant et al. (2004a) Sections 4.5 and 5.7. In particular, Beirlant et al. (1999) perform a joint

maximum likelihood for these three parameters at the same level k.

Another exponential approximation is ﬁrstly used in Feuerverger and Hall (1999). They

consider that for

γ>0

, the scaled log-spacings

, deﬁned in equation (12), are approximately

exponentially distributed with mean

γexp(β(n/i)ρ)

(with

β6=0

). They also proceed to the joint

maximum likelihood estimation of the three unknown parameters at the same level

. Considering

the same exponential approximation Gomes and Martins (2002) proposed a so-called external

estimation of the second order parameter

,i.e. its estimation at a level

higher than the level

used to estimate

, together with a ﬁrst order approximation for the maximum likelihood estimator

. They then obtain quasi-maximum likelihood explicit estimators of

and

, both computed

at the same level

, and through that external estimation of

. This reduces the asymptotic variance

of the

estimator comparatively to the asymptotic variance of the

estimator in Feuerverger and

Hall (1999), where

and

are estimated at the same level

.Gomes et al. (2007) build on

this approach and propose an external estimation of both

and

by maximum likelihood both

using a sample fraction

larger than the sample fraction

used to estimate

, also by maximum

likelihood. This reduces the bias without increasing the asymptotic variance, which is kept at the

value

γ2/k

, the asymptotic variance of Hill’s estimator. These estimators are thus better than the

Hill estimator for all k.

3.3. Control procedures

Extreme value theory and estimation in the distribution tail are greatly inﬂuenced by several

quantities. Firstly, we have to choose the tail sample fraction used for estimation. In this case,

procedures for optimal choice of this tail fraction are presented in Section 3.3.1. We can also use

graphical methods as presented in Section 3.3.2 to help to choose this tail fraction. Secondly, as

detailed in Section 2, the tail behaviour is very different depending on the value of the parameter

. Moreover, most of the estimates are not deﬁned for any

γ∈IR

but only for a smaller range

values. Some graphical procedures presented in Section 3.3.2 and the tests and conﬁdence

intervals presented in Section 3.3.3 can be used to assess the value of

, the domain of attraction

and the tail behaviour.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

90 M. Charras-Garrido and P. Lezaud

3.3.1. Optimal choice of the tail sample fraction

Practical application of the extreme value theory requires to select the tail sample fraction, i.e. the

extreme values of the sample that may contain most information on the tail behaviour. Indeed,

as illustrated in Figure 6for the Hill estimator, for a small tail sample fraction

, the

estimate

strongly differs when changing the value of

. Moreover, this estimation also greatly varies when

changing the sample for the same value of

, indicating a large variance of the estimate for small

values of

. Conversely, for large values of

, the

estimate presents a large bias, since the model

assumption may be strongly violated, but a smaller variance. Indeed, we observe in Figure 6that

for large values of k, the γestimates are close for the three simulated data sets.

FIG URE 6. Hill estimate of the extreme value index

against different values of

and three data sets of size

n=500

simulated from a Student distribution of parameter 3(with a true γ=1/3).

As noticed in Section 3.2.7, the bias of the estimates is controlled by the second order parame-

ters, including parameter

. These additional parameters have been used to propose estimators

with smaller bias and much less sensitive to changes in

. In the general case, the optimal

-value

depends on

and the parameters describing the second-order tail behaviour. Replacing these

second order parameters by their joint estimates yields an estimate for the optimal value of

. For

example, Guillou and Hall (2001) or Beirlant et al. (2004a) propose to choose the smallest value

of ksatisfying a given criterion which they deﬁned.

When the asymptotic mean and variance of the estimates are known, an important alternative is

to minimize the asymptotic mean squared error (AMSE) of the estimate of

, of a tail probability

or of a tail quantile, see e.g. Beirlant et al. (2004a). As detailed in the following Section, a mean

squared error plot representing the AMSE depending on the value of kcan also be useful.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 91

3.3.2. Graphical procedures

As noticed in Section 3.3.1, we need to select the tail sample fraction, e.g. the number of upper

extremes

, in order to apply extreme value theory for estimation purpose. Such a choice can be

supported visually by a diagram. To this aim estimates of

(see Figure 6), or other estimates, can

be plotted against different values of

. For small values of

, the variance of the estimator is large

and the bias is small, while for large values of

, the variance of the estimator is small and the

bias is large. In between, there is a balance between the variance and the bias and we observe a

plateau, where a suitable value of

may be chosen. Quite recent estimators, see e.g. Section 3.2.7,

have the interesting property to present a relatively large plateau, that makes the choice of an

appropriate value of

, less critical. To explore this balance between the variance and the bias,

another option consists in plotting against the value of

, a mean square error computed, either

from the true value when studying an estimate with simulated data sets or from an estimation

obtained from real data sets.

As noticed above, the estimates of the extreme value index

, and consequently the tail

estimation, can be very different depending on the selected tail sample fraction. In particular, for

large values of

, the model assumption may be strongly violated. It is then important to check the

validity of the model. Thus, we present some graphical assessments for the validity of extreme

value extrapolation. Firstly, we can use a probability plot (or PP-plot) which is a comparison of

the empirical and ﬁtted distribution functions, that may be equivalent if the model is convenient.

For example, considering the ordered block maximum data

Z(1)≥.. . ≥Z(m)

, the PP-plot will

consist of the points

m+1,Gˆ

γm,ˆ

µm,ˆ

σm(Z(i)) = exp −1+ˆ

γm

Z(i)−ˆ

µm

σm−1/ˆ

γm!!for i=1,...,m.

We can also draw the PP-plot with the original sample. For example, in the POT case, we can

represent the points (see Figure 7) as follows

i

,1−kn

nHˆ

γkn,ˆ

σkn(Xn−kn+i+1,n−Xn−kn+1,n)for i=1,...,kn.

Secondly, we can use a quantile plot (or QQ-plot) which is a comparison between the empirical

and model estimated quantiles, that may also be equivalent if the model is convenient. For example,

the ordered block maximum data lead to plot the points

G−1

γm,ˆ

µm,ˆ

σmi

m+1=ˆ

µm+ˆ

σm

γm1−−log i

m+1−ˆ

γm!for i=1,...,m.

Again, we can also draw the QQ-plot with the original sample. For example, in the POT case, we

can represent the points (see Figure 8)

Xn−kn+i+1,n,H−1

γkn,ˆ

σknkn

n1−i

n+Xn−kn+1,nfor i=1,...,kn.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

92 M. Charras-Garrido and P. Lezaud

FIG URE 7. Example of PP-plot for the original sample.

In all the above mentioned PP or QQ-plots, the points should lie close to the unit diagonal.

Substantial departures from linearity lead to suspect that either the parameter estimation method

or the selected model (related for example to the chosen tail sample fraction) is inaccurate. A

weakness of the PP-plot is that there is an over-smoothing, particularly in the upper and the lower

tails of the distribution. Especially, the both coordinates are bounded to

for the largest data,

i.e. the one of greatest interest for extreme values. Then, the probability plot provides the least

information in the region of most interest. In consequence, Reiss and Thomas (2007) recommend

to use the PP-plot principally to justify an hypothesis visually. They suggest to use other tools,

including QQ-plot, whenever a critical attitude towards modelling is adopted. Indeed, a QQ-plot

achieves a better compromise between the reduction of random data ﬂuctuations and exhibition of

special features and clues contained in the data.

There exists several other graphical tools including return level plots whose principles are

analogous to those of the PP and QQ-plots. The density plot compares the density estimated by

the model to a non-parametric estimation, e.g. histogram or kernel estimate. They are mainly

of interest when the goal is to produce an estimation of the distribution tail and are not used

when the goal is to estimate the extreme value index

. Different variants of PP-plot or QQ-plot

include a log-transform of the coordinates of the points. For example, the Hill and Zipf estimates

(see Sections 3.2.3 and 3.2.4) are based on a generalized quantile plot. We will now focus in

particular on the Gumbel plot. It is based on the fact that in the Gumbel maximal domain of

attraction the excesses are exponentially distributed with parameter 1. The Gumbel plot consists

in plotting the quantiles

−log(i/k)

against the ordered excesses

Xn−k+i,n−Xn−k,n

as in Figure 9.

In the Gumbel domain of attraction, see left panel of Figure 9, the points should lie close to the

unit diagonal and the slope of the graph will give an estimate of the shape parameter, e.g.

for

the GPD. In the Fréchet domain of attraction an upward curvature may appear (see central panel

of Figure 9), while a downward curvature may indicate a Reversed Weibull domain of attraction

(see left panel of Figure 9). Outliers may also be detected using this plot. This last plot is mainly

used to graphically assess the domain of attration of a data set.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 93

FIG URE 8. Example of QQ-plot for the original sample.

3.3.3. Tests and Conﬁdence intervals

For many estimates, e.g. maximum likelihood or probability weighted moments, approximate

normality is established, and conﬁdence intervals for the GEV (or GPD) parameters follow as

detailed for example in Castillo et al. (2004) Section 9.2. Direct application of the delta method

yields approximate normality for the quantile corresponding estimates, and conﬁdence intervals

for the quantile can be deduced, as presented e.g. in Castillo et al. (2004). In other cases, the

variance of the estimates may not be analytically readily available. In such cases, an estimate of

the variance can be obtained using sampling methods such as jackknife and bootstrap methods

presented in Efron (1979), with a preference for parametric bootstrap. In this simulation context,

conﬁdence intervals are obtained selecting empirical quantiles from the estimates (of parameters

or quantiles) computed on a large number of simulated samples.

GEV has three special cases that have very different tail behaviours. For example, a distribution

with a ﬁnite endpoint (

ω(F)

) cannot be in the Fréchet domain of attraction, and conversely an

unlimited distribution cannot be in the Reversed Weibull domain of attraction. Moreover, many

estimates are limited to some ranges of the extreme value index

. Model selection then focuses

on deciding which one of these GEV particular case best ﬁts the data. In particular, we wish

to test

H0:γ=0

(Gumbel) versus

H1:γ6=0

(Fréchet or Weibull), or

H1:γ<0

(Weilbull), or

H1:γ>0

(Fréchet). To this end, we can estimate

for the GEV (or GPD) model using the

maximum likelihood and perform a likelihood ratio test as detailed for example in Castillo et al.

(2004) Sections 6.2 and 9.6. We can also use a conﬁdence interval for

, then check if it contains

the value 0 and ﬁnally decide accordingly.

4. Conclusion

In this article, we presented the probability framework and the statistical analysis of extreme

values. The probability framework starts with the famous Fisher-Tippett-Gnedenko Theorem 2.3

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

94 M. Charras-Garrido and P. Lezaud

FIG URE 9. Gumbel plot for the Gumbel (left panel), Fréchet (central panel) and Reversed Weibull (right panel) domains

of attraction.

which characterizes the three types of max-stable distributions. It remains to ﬁnd the necessary

and sufﬁcient conditions to determine the domain of attraction of a speciﬁc distribution. The

main tool to address this question is the notion of regular variation which plays an essential role

in many limit theorems. Moreover, the Fisher-Tippett-Gnedenko Theorem restricts itself to iid

random variables, so the necessity to modify the standard approach for analysing the extremes

of stationary time series for instance. The results mainly obtained by M. R. Leadbetter ends the

probability part of this article. We deliberately limited our presentation to the bases of the theory,

so the point process approach has just been alluded and we omitted the multivariate extremes (see

chapter 8 in Beirlant et al. (2004b)). The exceedances of a stochastic process, i.e. the study of

P(max0≤s≤tXs≥b)

for a stochastic process

(Xt)

are addressed in Aldous (1989), Berman (1992)

and Falk et al. (2011). In addition, Adler (2000) and Azaïs and Wschebor (2009) are mainly

dedicated to level sets and extrema of Gaussian random ﬁelds. At the heart of the Adler’s approach

stands the use of the Euler characteristic of level sets, whereas the book of Azaïs and Wschebor

relies on Rice formula, a general tool to get results on the moments of the number of crossings of

a given level by the process.

Estimating the distribution tail is a difﬁcult problem since it implies an extrapolation. As a

sign of this difﬁculty, numerous estimators have been proposed, some of them very recently, and

none of them have made consensus. According to the application (and so the expected value of

the customs and practices of the applied ﬁeld, the quantities of interest (estimation of

, of the

distribution tail, or of an extreme quantile) or the expected properties (low sensitivity to changes in

, low bias, low variance) different estimators can be chosen. The choice of an estimator can also

be driven by practical considerations, since only some of the estimates proposed in the literature

are available in classical softwares. A recent list can be ﬁnd in Gilleland et al. (2013) and can

help to choose estimates that are already implemented and then easy to apply. Extreme value

modelling is still an active ﬁeld. Topics like threshold or tail sample fraction selection, trends and

change points in the tail behaviour, clustering, rates of convergence or penultimate approximations,

among others, are still challenging. More details on open research topics concerning univariate

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 95

extremes are given by Beirlant et al. (2012). Other challenges concern spatial extremes or non iid

observations. Bases on spatial extremes can be found e.g. in Falk et al. (2011) or Castillo et al.

(2004). Elements on extreme analysis of non iid observations are presented in Falk et al. (2011).

The statistical analysis of extreme values needs a long observation time because of the very low

probability of the events considered. In many applications, such as complex systems with many

interactions, collecting data is difﬁcult, if not impossible. An alternative approach consists in the

modelling of the process leading to the feared event. To achieve this, the ﬁrst step requires that the

considered system is formalized and only then, some estimate can be obtained by using simulation

tools. Nevertheless, obtaining accurate estimates of rare event probabilities using traditional

Monte Carlo techniques requires a huge amount of computing time. Many techniques for reducing

the number of trials in Monte Carlo simulation have been proposed, the more promising is based

on importance sampling. But to use importance sampling, we need to have a deep knowledge of

the studied system and, even in such a case, importance sampling may not provide any speed-up.

An alternative way to increase the relative number of visits to the rare event is to use trajectory

splitting, based on the idea that there exist some well identiﬁable intermediate system states that

are visited much more often than the target states themselves and behave as gateway states to

reach the target states. For more details of the simulation of rare events, we suggest consulting

Doucet et al. (2001), Bucklew (2011) and Rubino and Tufﬁn (2009).

Acknowledgement

We sincerely thank the Associate Editor and the referees for their careful reading, constructive

comments, and relevant remarks.

References

Adler, R. J. (2000). On excursion sets, tube formulas and maxima of random ﬁelds. Ann. Appl. Probab., 10(1):1–74.

Aldous, D. (1989). Probability Approximations via the Poisson Clumping Heuristic. Springer Verlag.

Azaïs, J. M. and Wschebor, M. (2009). Level sets and extrema of random processes and ﬁelds. John Wiley & Sons.

Beirlant, J., Caeiro, F., and Gomes, M. (2012). An overview and open research topics in statistics of univariate extremes.

REVSTAT - Statistical Journal, 10(1):1–31.

Beirlant, J., Dierckx, G., Goegebeur, Y., and Matthys, G. (1999). Tail index estimation and an exponential regression

model. Extremes, 2(2):177–200.

Beirlant, J., Dierckx, G., Guillou, A., and Sta˘

arica˘

a, C. (2002). On exponential representations of log-spacings of

extreme order statistics. Extremes, 5(2):157–180.

Beirlant, J., Goegebeur, Y., Segers, J., and Teugels, J. (2004a). Statistics of Extremes. Wiley Series in Probability and

Statistics. John Wiley & Sons, Ltd.

Beirlant, J., Goegebeur, Y., Segers, J., and Teugels, J. (2004b). Statistics of Extremes: Theory and Applications.

Probability and Statistics. Wiley.

Berman, S. M. (1992). Sojourns and extremes of stochastic processes. Wadsworth and Brooks.

Billingsley, P. (1995). Probability and Measure. Wiley, 3ème edition.

Bingham, N., Goldie, C. M., and Teugels, J. L. (1989). Regular Variation. Cambridge University Press.

Bouleau, N. (1991). Splendeurs et misères des lois de valeurs extrêmes. Extremes.

http://halshs.

archives-ouvertes.fr/docs/00/05/65/72/PDF/c15.pdf.

Bucklew, J. (2011). Introduction to rare event simulation. Springer Series in Statistics. Springer-Verlag.

Castillo, E., Hadi, A., Balakrishnan, N., and Sarabia, J. (2004). Extreme value and related models with applications in

engineering and science. Wiley.

Coles, S. (2001). An introduction to statistical modeling of extreme values. Springer.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

96 M. Charras-Garrido and P. Lezaud

Coles, S. and Powell, E. (1996). Bayesian methods in extreme value modelling: A review and new developments.

International Statistical Review, 64(1):119–136.

Csorgo, S. and Viharos, L. (1998). Estimating the tail index, pages 833–881.

Davis, R. (1982). Limit laws for the maximum and minimum of stationary sequences. Zeitschrift für Wahrschein-

lichkeitstheorie und Werwandtl Gebiete,61:31–42.

de Haan, L. and Ferreira, A. (2006). Extreme value theory: An introduction. Springer.

Dekkers, A. and de Haan, L. (1989). On the estimation of the extreme-value index and large quantile estimation.

Annals of Statistics, 17(4):1795–1832.

Dekkers, A., Einmahl, J., and de Haan, L. (1989). A moment estimator for the index of an extreme-value distribution.

Annals of Statistics, 17(4):1833–1855.

Denzel, G. and O’Brien, G. (1975). Limit theorems for extreme values of chain-dependent processes. Annals of

Probability, 3:773–779.

Diebolt, J., El-Aroui, M., Garrido, M., and Girard, S. (2005). Quasi-conjugate bayes estimates for gpd parameters and

application to heavy tails modelling. Extremes, 8(1-2):57–78.

Diebolt, J., Guillou, A., Naveau, P., and Ribereau, P. (2008). Improving probability-weighted moment methods for the

generalized extreme value distribution. REVSTAT - Statistical Journal, 6(1):33–50.

Diebolt, J., Guillou, A., and Rached, I. (2007). Approximation of the distribution of excesses through a generalized

probability-weighted moments method. Journal of Statistical Planning and Inference, 137(3):841–857.

do Nascimento, F., Gamerman, D., and Lopes, H. (2011). A semiparametric bayesian approach to extreme value

estimation. Statistics and Computing, 22(2):661–675.

Doucet, A., de Freitas, N., and Gordon, N. (2001). An Introduction to Sequential Monte Carlo Methods. Springer

Verlag.

Efron, B. (1979). Bootstrap methods: another look at the jackknife. The annals of Statistics, 7(7):1–26.

Embrechts, P., Klüppelberg, C., and Mikosch, T. (2003). Modelling Extremal Events for Insurance and Finance,

volume 33 of Applications of Mathematics. Springer.

Falk, M., Hüsler, J., and Reiss, R. D. (2011). Laws of small numbers: extremes and rare events. Birkhäuser/Springer.

Feller, W. (1971). An Introduction to Probability Theory and Its Applications II. Wiley & Sons, 2ème edition.

Feuerverger, A. and Hall, P. (1999). Estimating a tail exponent by modelling departure from a pareto distribution.

Annals of Statistics, 27(2):760–781.

Fisher, R. and Tippett, L. (1928). Limiting forms of the frequency distribution of the largest and smallest member of a

sample. Proc. Cambridge Phil. Soc., 24:180–190.

Fraga Alves, M. (2002). A location invariant hill-type estimator. Extremes, 4(3):199–217.

Gilleland, E., Ribatet, M., and Stephenson, A. (2013). A software review for extreme value analysis. Extremes,

16(1):103–119.

Gomes, M. and Martins, M. (2002). “asymptotically unbiased” estimators of the tail index based on external estimation

of the second order parameter. Extremes, 5(1):5–31.

Gomes, M., Martins, M., and Neves, M. (2007). Improving second order reduced bias extreme value index estimation.

REVSTAT - Statistical Journal, 5(2):177–207.

Greenwood, J., Landwehr, J., Matalas, N., and Wallis, J. (1979). Probability weighted moments: Deﬁnition and relation

to parameters of several distributions expressable in inverse form. Water Ressources Research, 15(5):1049–1054.

Guillou, A. and Hall, P. (2001). A diagnostic for selecting the threshold in extreme value analysis. Journal of the

Royal Statistical Society: Series B (Statistical Methodology), 63(2):293–305.

http://doi.wiley.com/10.1111/

1467-9868.00286.

Hill, B. (1975). A simple general approach to inference about the tail of a distribution. The annals of statistics,

3(5):1163–1174.

Hosking, J. (2013). Algorithm as 215: Maximum-likelihood estimation of the parameters of the generalized extreme-

value distribution. Journal of the Royal Statistical Society: Series C (Applied Statistics), 34(3):301–310.

Hosking, J. and Wallis, J. (1987). Parameter and quantile estimation for the generalized pareto distribution. Technomet-

rics, 29(3):339–349.

Karamata, J. (1933). Sur un mode de croissance régulière. Théorèmes fondamentaux. Bull. Soc. Math. France,

61:55–62.

Korevaar, J. (2004). Tauberian Theory: a century of developments, volume 329 of A Series of Comprehensive Studies

in Mathematics. Springer.

Leadbetter, M. (1974). On extreme values in stationary sequences. Zeitschrift für Wahrscheinlichkeitstheorie und

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

Extreme Value Analysis: an Introduction 97

Werwandtl Gebiete,28:289–303.

Leadbetter, M. (1983). Extremes and local dependence in stationary sequences. Zeitschrift für Wahrscheinlichkeitsthe-

orie und Werwandtl Gebiete,65:291–306.

Leadbetter, M., Lindgren, G., and Rootzén, H. (1983). Extremes and Related Properties of Random Sequences and

Processes. Springer-Verlag.

Macleod, A. (1989). Algorithm as 245: A robust and reliable algorithm for the logarithm of the gamma function.

Journal of the Royal Statistical Society: Series C (Applied Statistics), 38(2):397–402.

Mason, D. (1982). Laws of large numbers for sums of extreme values. The Annals of Probability, 10(3):754–764.

O’Brien, G. (1974). The maximum term of uniformly mixing stationary processes. Zeitschrift für Wahrscheinlichkeits-

theorie und Werwandtl Gebiete,30:57–63.

O’Brien, G. (1987). Extreme values for stationary and markov sequences. Annals of Probability,15:281–291.

Pickands, J. (1975). Statistical inference using extreme order statistics. Annals of Statistics, 3(1):119–131.

Prescott, P. and Walden, A. (1980). Maximum likelihood estimation of the parameters of the generalized extreme-value

distribution. Biometrika, 67(3):723–724.

Prescott, P. and Walden, A. (1983). Maximum likelihood estimation of the parameters of the three-parameter

generalized extreme-value distribution from censored samples. Journal of Statistical Computation and Simulation,

16(3-4):241–250.

Reiss, R. D. and Thomas, M. (1997). Statistical Analysis of Extreme Values. Birkhäuser.

Reiss, R.-D. and Thomas, M. (2007). Statistical analysis of extreme values: with applications to insurance, ﬁnance,

hydrology and other ﬁelds. Birkhäuser, third edition.

Resnick, S. (1987). Extreme Values, Regular Variation, and Point Processes. Springer Verlag.

Rubino, G. and Tufﬁn, B., editors (2009). Rare event simulation using Monte Carlo methods. John Wiley & Sons.

Segers, J. (2005). Generalized pickands estimators for the extreme value index. Journal of Statistical Planning and

Inference, 128(2):381–396.

Smith, R. (1985). Maximum likelihood estimation in a class of nonregular cases. Biometrika, 72(1):67–90.

Stephenson, A. and Tawn, J. (2004). Bayesian inference for extremes: Accounting for the three extremal types.

Extremes, 7(4):291–307.

von Mises, R. (1975). La distribution de la plus grandes de

valeurs. Reprinted(1954) in Selected Papers II: Amer.

Math. Soc., Providence, R.I., pages 271–294.

Zhou, C. (2009). Existence and consistency of the maximum likelihood estimator for the extreme value index. Journal

of Multivariate Analysis, 100(4):794–815.

Zhou, C. (2010). The extent of the maximum likelihood estimator for the extreme value index. Journal of Multivariate

Analysis, 101(4):971–983.

Journal de la Société Française de Statistique,Vol. 154 No. 2 66-97

http://www.sfds.asso.fr/journal

The distribution of fitness effects during adaptive walks using a simple genetic network

Article

Full-text available

May 2024
PLOS GENET

The tempo and mode of adaptation depends on the availability of beneficial alleles. Genetic interactions arising from gene networks can restrict this availability. However, the extent to which networks affect adaptation remains largely unknown. Current models of evolution consider additive genotype-phenotype relationships while often ignoring the contribution of gene interactions to phenotypic variance. In this study, we model a quantitative trait as the product of a simple gene regulatory network, the negative autoregulation motif. Using forward-time genetic simulations, we measure adaptive walks towards a phenotypic optimum in both additive and network models. A key expectation from adaptive walk theory is that the distribution of fitness effects of new beneficial mutations is exponential. We found that both models instead harbored distributions with fewer large-effect beneficial alleles than expected. The network model also had a complex and bimodal distribution of fitness effects among all mutations, with a considerable density at deleterious selection coefficients. This behavior is reminiscent of the cost of complexity, where correlations among traits constrain adaptation. Our results suggest that the interactions emerging from genetic networks can generate complex and multimodal distributions of fitness effects.

Comparison of Future Design Rainfall with Current Design Rainfall: A Case Study in New South Wales, Australia

Article

Full-text available

Jun 2024

Climate change impacts have the potential to alter the design rainfall estimates around the world. Decreasing trends in the summer and winter rainfall in New South Wales (NSW), Australia have already been observed due to climate variability and change. The derivation of design rainfall from historical rainfall, which is required for the design of stormwater management infrastructure, may be ineffective and costly. It is essential to consider climate change impacts in estimating design rainfall for the successful design of stormwater management infrastructure. In this study, the probability of the occurrence of daily extreme rainfall has been assessed under climate change conditions. The assessment was performed using data from 29 meteorological stations in NSW, Australia. For the evaluation of future design rainfall, the probability of the occurrence of extreme rainfall for different recurrence intervals was developed from daily extreme rainfall for the periods of 2020 to 2099 and compared with the current Australian Bureau of Meteorology (BoM) design rainfall estimates. The historical mean extreme rainfall across NSW varied from 37.71 mm to 147.3 mm, indicating the topographic and climatic influences on extreme rainfall. The outcomes of the study suggested that the future design rainfall will be significantly different from the current BoM estimates for most of the studied stations. The comparison of the results showed that future rainfall in NSW will change from −4.7% to +60% for a 100-year recurrence interval. However, for a 2-year recurrence interval, the potential design rainfall change varies from an approximately 8% increase to a 40% decrease. This study revealed that the currently designed stormwater management infrastructure will be idle in the changing climate.

FINANCIAL RISK MANAGEMENT AND FINANCIAL PERFORMANCE OF LISTED DEPOSIT MONEY BANKS IN NIGERIA

Article

Full-text available

Jan 2024

Doutimiareye Newstyle Samuel

Distinguishing subsampled power laws from other heavy-tailed distributions

Article

Full-text available

May 2024
PRE

Distinguishing power-law distributions from other heavy-tailed distributions is challenging, and this task is often further complicated by subsampling effects. In this work, we evaluate the performance of two commonly used methods for detecting power-law distributions—the maximum likelihood method of Clauset and the extreme value method of Voitalov —in distinguishing subsampled power laws from two other heavy-tailed distributions, the lognormal and the stretched exponential distributions. We focus on a random subsampling method commonly applied in network science and biological sciences. In this subsampling scheme, we are ultimately interested in the frequency distribution of elements with a certain number of constituent parts—for example, species with k individuals or nodes with k connections—and each part is selected to the subsample with an equal probability. We investigate how well the results obtained from low-subsampling-depth subsamples generalize to the original distribution. Our results show that the power-law exponent of the original distribution can be estimated fairly accurately from subsamples, but classifying the distribution correctly is more challenging. The maximum likelihood method falsely rejects the power-law hypothesis for a large fraction of subsamples from power-law distributions. While the extreme value method correctly recognizes subsampled power-law distributions with all tested subsampling depths, its capacity to distinguish power laws from the heavy-tailed alternatives is limited. However, these false positives tend to result not from the subsampling itself but from the estimators' inability to classify the original sample correctly. In fact, we show that the extreme value method can sometimes be expected to perform better on subsamples than on the original samples from the lognormal and the stretched exponential distributions, while the contrary is true for the main tests included in the maximum likelihood method. ©2024 American Physical Society 2024 American Physical Society

Star entrepreneurs on digital platforms: Heavy-tailed performance distributions and their generative mechanisms

Article

Jan 2024
J BUS VENTURING

This study extends emerging theories of star performers to digital platforms, an increasingly prevalent entrepreneurial context. It hypothesizes that the unique characteristics of many digital platforms (e.g., low marginal costs, feedback loops, and network effects) produce heavy-tailed performance distributions, indicating the existence of star entrepreneurs. Using longitudinal data from an online learning platform, proportional differentiation is identified as the most likely generative mechanism and lognormal distribution as the most likely shape for distributions of entrepreneurial performance in digital contexts. This study contributes theory and empirical evidence for non-normal entrepreneurial performance with implications for scholars and practitioners of digital entrepreneurship. Executive summary The performance of 'star' entrepreneurs on digital platforms can be 100-or 1000-fold that of their average competitors. When performance is plotted as a distribution, star performers reside in the tails of these distributions. The assumption of a normal distribution of performance in the bulk of entrepreneurship research implies that most performance observations are clustered around the average. Instead, most entrepreneurs on digital platforms exhibit sub-par performance, while a minority captures a major fraction of the generated value. This paper argues that the unique characteristics of digital contexts-nearly zero marginal costs, feedback loops, and network effects-drive such extreme performance. Using data from Udemy, a digital platform where independent producers (entrepreneurs) offer educational videos (digital products) to a large pool of potential customers, we provide evidence that entre-preneurial performance is lognormally rather than normally distributed. We further identify proportional differentiation as the underlying generative mechanism. Thus, star performance on digital platforms is not driven only by the rich-get-richer effect. Instead, both the initial value of performance and the rate at which it is accumulated play important roles in explaining extreme performance outcomes. This discovery has important implications for entrepreneurship theory and practice. Our findings, for example, signal that some late entrants who successfully pursue high customer accumulation rates in domains with high knowledge intensity can become star entrepreneurs.

More than Magnitude: Towards a multidimensional understanding of unprecedented weather to better support disaster management

Article

Dec 2024

Comparative Analysis of Outlier Detection Approaches in MetOcean Domain: Extreme Value Analysis, Functional Data Analysis, and Unsupervised Learning

Preprint

Full-text available

Jun 2024

Detecting outliers within MetOcean datasets is crucial for identifying extreme weather conditions to facilitate informed decision-making whilst various outlier detection approaches and methods are available to be selected. However, choosing the most appropriate one requires an in-depth understanding of their capabilities and this poses a great challenge to the novice within the domain. Therefore, this study is proposed to empirically compare these approaches and provide insights on their performance when dealing with large MetOcean data. The comparison is driven by three outlier detection approaches using the SEAFINE dataset: Extreme Value Analysis (EVA), Functional Data Analysis (FDA), and Unsupervised Learning (UL). A framework is designed to systematically conduct the comparative analysis, which involves five phases: data preprocessing, model training, outlier detection, parameter estimation, and evaluation. This analysis provides insights into the performance of each outlier detection method for each approach and the comparison between them. Overall, the results indicate that the EVA approach exhibits the fastest average execution time compared to the FDA and UL. Within the EVA approach, Peak-Over-Threshold is the most effective, closely aligning the distribution of extreme values with the estimated distribution. Functional Boxplot emerges as the optimal method for FDA, capable of detecting outliers far from the median. Cluster-Based Local Outlier Factor (CBLOF) is the most effective UL method, predicting outliers close to the expected total, with high Excess-Mass values and fast execution.

Thermodynamic state shift observed prior to the 2011 Mw 9 East Japan earthquake through data mining using cellular automaton

Article

Full-text available

Jan 2024

Hiroyuki Kikuchi

In this study, we employed our previously developed data mining method to show that a thermodynamic state shift occurred preceding the 2011 Mw 9 East Japan Earthquake (GEJE), coinciding with the onset of crustal stress manifestations. Our discussion starts with the insights obtained from our prior research, which revealed that small ground vibration fluctuations (GVF) detected near the epicenter of the GEJE are thermodynamically equivalent to a cellular automaton (CA). This equivalence allowed us to consider the “Process to GEJE,” defined as the transition from GVF to catastrophic rupture in the GEJE, as analogous to the nonequilibrium-to-equilibrium thermodynamic transition in the CA. Therefore, if we find a thermodynamic state shift in the CA, it is reasonable to assume that a similar thermodynamic state shift may exist in the “Process to GEJE.” We successfully derived a thermodynamic state function that exhibits a significant change before the nonequilibrium-to-equilibrium thermodynamic transition in the CA. By evaluating this thermodynamic state function using the GVF data of the “Process to GEJE,” we discovered its maximum value occurring on 2007-10-28 at 10:30:00 JST, 1231 days before the GEJE occurred on 2011-03-11 at 14:46 JST. This maximum value can be considered as an indication of a thermodynamic state shift prior to the GEJE. Furthermore, near the timing of the thermodynamic state shift, it has been observed that the GVF state approaches thermodynamic equilibrium and a sudden increase in crustal stress was first noted on 2007-11-22 at 18:02:44 JST.

A Non-stationary Stochastic Model of Extreme Rain Events in the Changing Climate

Chapter

Aug 2023

The Intergovernmental Panel on Climate Change has concluded that the frequency and intensity of extremes of heat waves, precipitation, droughts, and cyclones will continue to increase with every additional increment of global warming. Time-dependent changes in intensity and frequency of weather extremes require the non-stationary stochastic modelling of various environmental loads in the structural reliability assessment of infrastructure systems. Since most of the existing design codes and standards assume ‘stationary climate’ conditions, the effects of the transition to non-stationary conditions on infrastructure risk and reliability assessment must be carefully investigated. Whether the current design philosophies will ensure proper safety for our structures under these changing conditions, is a matter of concern from the engineering perspective. Therefore, this paper develops a non-stationary Poisson process model of extreme rain events to support the assessment of the increased risk of flooding in the future. The proposed model has a baseline model of precipitation events that are based on historical data available from existing weather stations. Then, time-dependent amplification functions, based on the findings of climate models, are assigned to the frequency and intensity of rain events. To illustrate the proposed approach, the paper analyses precipitation data from Toronto International Airport Station. The results indicate that the return periods of the extreme one-day rainfall events in Toronto corresponding to higher threshold magnitudes decrease significantly over time. The distribution of the extreme values of one-day rainfall events increases considerably with increase in design life of the structures.KeywordsExtreme rain eventsNon-stationary stochastic model

BER analysis of an optimum MIMO linear receiver in optical SDM systems with mode-dependent loss

Article

Full-text available

Sep 2023
OPT EXPRESS

Multiple-input multiple-output (MIMO) receivers designed to optimize the minimum mean square error (MMSE) are a common choice in coherent optical communication systems based on spatial division multiplexing (SDM). This kind of receivers naturally integrate both MIMO equalization and matched filtering functions. However, when the optical channel exhibits significant mode-dependent loss (MDL) and/or mode-dependent gain (MDG), the impact of inter-symbol interference (ISI) and crosstalk that arise, even using an ideal MIMO MMSE linear receiver, is barely analyzed. Moreover, due to the random nature of the MDL/MDG model, the resulting ISI, crosstalk, and bit error rate (BER) also become random variables and, hence the system performance is more unpredictable. In this paper, we first evaluate the residual distortion (ISI and crosstalk) after the MIMO receiver and then we study the validity of assuming it as an additional Gaussian noise term independent of the channel noise. Next, the probability density distribution (PDF) of the BER is analyzed, from both an analytical perspective and numerical simulations. For the latter, we use a single-carrier 2-PAM (pulse amplitude modulation) system, with pulse shaping, and the MIMO MMSE receiver implementation by means of a MIMO fractionally-spaced equalizer (FSE). We carry out simulations of the system under different conditions of MDL/MDG level and signal to noise ratio (SNR), measured at the receiver input. Additionally, we address possible fits of the BER PDF to known closed-form distributions, among which the Generalized Extreme Value (GEV) family of distributions is selected, and polynomial functions are proposed that relate the system parameters with the GEV PDF parameters. Finally, we present contour maps of BER according to a giving target of system outage probability (OP) that depend on the MDL/MDG and SNR conditions.

An introduction to probability theory and its applications, Vol. II

Article

May 1967
J FRANKLIN I

H.C. Ko

Sur un mode de croissance régulière. Théorèmes fondamentaux

Article

J. Karamata

MAXIMUM-LIKELIHOOD ESTIMATION IN A CLASS OF NONREGULAR CASES

Article

Jan 1985

RL SMITH

An Introduction To Probability Theory And Its Applications

Article

Jun 1958

Book

Jan 1983

Probability and Measure

Book

Jan 1986
J AM STAT ASSOC

P. Billingsley

Probability. Measure. Integration. Random Variables and Expected Values. Convergence of Distributions. Derivatives and Conditional Probability. Stochastic Processes. Appendix. Notes on the Problems. Bibliography. List of Symbols. Index.

Extreme Value and Related Models With Applications in Engineering and Science

Book

Jan 2005

Extreme values for stationary sequences

Article

Jan 1986

George L. O’Brien

Rare Event Simulation Using Monte Carlo Methods

Article

Mar 2009

In a probabilistic model, a rare event is an event with a very small probability of occurrence. The forecasting of rare events is a formidable task but is important in many areas. For instance a catastrophic failure in a transport system or in a nuclear power plant, the failure of an information processing system in a bank, or in the communication network of a group of banks, leading to financial losses. Being able to evaluate the probability of rare events is therefore a critical issue. Monte Carlo Methods, the simulation of corresponding models, are used to analyze rare events. This book sets out to present the mathematical tools available for the efficient simulation of rare events. Importance sampling and splitting are presented along with an exposition of how to apply these tools to a variety of fields ranging from performance and dependability evaluation of complex systems, typically in computer science or in telecommunications, to chemical reaction analysis in biology or particle transport in physics. Graduate students, researchers and practitioners who wish to learn and apply rare event simulation techniques will find this book beneficial.

Statistical analysis of extreme values. With applications to insurance, finance, hydrology and other fields. With CD-ROM

Article

Jan 2007

The statistical analysis of extreme data is important for various disciplines, including hydrology, insurance, finance, engineering and environmental sciences. This book provides a self-contained introduction to the parametric modeling, exploratory analysis and statistical interference for extreme values. The entire text of this third edition has been thoroughly updated and rearranged to meet the new requirements. Additional sections and chapters, elaborated on more than 100 pages, are particularly concerned with topics like dependencies, the conditional analysis and the multivariate modeling of extreme data. Parts I-III about the basic extreme value methodology remain unchanged to some larger extent, yet notable are, e.g., the new sections about "An Overview of Reduced-Bias Estimation" (co-authored by M.I. Gomes), "The Spectral Decomposition Methodology", and "About Tail Independence" (co-authored by M. Frick), and the new chapter about "Extreme Value Statistics of Dependent Random Variables" (co-authored by H. Drees). Other new topics, e.g., a chapter about "Environmental Sciences", (co-authored by R.W. Katz), are collected within Parts IV-VI.

Extreme Value Analysis : an Introduction

Figures

Recommended publications

INRAE (France) is recruiting 55 research scientists

INRAE is hiring 10 research scientists - Call for research projects (CRCN)

INRAE is hiring 45 Scientists through open competitions and offering permanent positions.

Arethusa: Introduction

Etudes Anglaises: Introduction

Introduction: Profile of an eco-literature

Social Research: Introduction

Quaternions in computer graphics and robotics