Extreme value theory as a risk management tool

Paul Embrechts, Sidney I. Resnick and Gennady Samorodnitsky

Abstract
The nancial industry, including banking and insurance, is undergoing ma-
jor changes. The (re)insurance industry is increasingly exposed to catastroph-
ic losses for which the requested cover is only just available. Due to an in-
creasing complexity of nancial instruments, sophisticated risk management
tools have to be put into place. The securitization of risk and alternative risk
transfer highlight the convergence of nance and insurance at the product
level. Extreme value theory plays an important methodological role within
the above.

1 Introduction
Consider the time series in Table 1 of loss{ratios (yearly data) for earthquake insur-
ance in California from 1971 till 1993. The data are taken from Jae and Russell
(1996).
1971 17.4 1979 2.2 1987 22.8
1972 0 1980 9.2 1988 11.5
1973 0.6 1981 0.9 1989 129.8
1974 3.4 1982 0 1990 47.0
1975 0 1983 2.9 1991 17.2
1976 0 1984 5.0 1992 12.8
1977 0.7 1985 1.3 1993 3.2
1978 1.5 1986 9.3

Table 1: California earthquake data

 A rst version of this paper was presented by the rst author as an invited paper at the
XXVIIIth International ASTIN Colloquium in Cairns and published in the conference proceedings
under the title \Extremes and Insurance". Sidney Resnick and Gennady Samorodnitsky were
partially supported by NSF Grant DMS-97-04982 at Cornell University. Paul Embrechts gratefully
acknowledges the hospitality of the School of Operations Research and Industrial Engineering,
Cornell University during Fall 1996.

1
 2
On the basis of these data, who would have guessed the 1994 value of 2272.7? In-
deed, on the 17th of January of that year the 6.6 Richter scale Northridge earthquake
hit California causing an insured damage of USD 10.4 billion and a total damage of
USD 30 billion, making 1994 the year with the third highest loss burden (natural
catastrophes and major losses) in the history of insurance. The front-runners in this
sad hit parade are 1992 (the year of hurricane Andrew) and 1990 (the year of the
winter storms Daria and Vivian). For details on these, see Sigma (1995, 1997).
The reinsurance industry experienced a rise in both intensity as well as magnitude
of losses due to natural and man{made catastrophes. For the United States alone,
Canter, Cole and Sandor (1996) estimate an approximate USD 245 billion of capital
in the insurance and reinsurance industry to service a country that has USD 25
to 30 trillion worth of property. No surprise that the nance industry has seized
upon this by oering (often in joint ventures with the (re)insurance world) properly
securitized products in the realm of catastrophe insurance. At an increasing pace,
new products are being born. Some of them only have a short life, and others are
reborn under a dierent shape. Some do not survive. Examples include:
| CAT futures and PCS options (CBOT): in these cases, securitization is achiev-
ed through the construction of derivatives written on a newly constructed
industry wide loss{ratio index.
| Convertible CAT bonds. The Winterthur convertible hail{bond is an example.
This European type convertible has an extra coupon payment contingent on
the occurrence of a well{dened cat{event: an excessive number of cars in
Winterthur's Swiss portfolio damaged within a hail storm over a specic time
period. For details, see Schmock (1997).
Further interesting new products are the so{called multi{line, multi{year, high layer
(infrequent event) products, credit lines, the catastrophe risk exchange (CATEX).
For a brief review on some of these instruments, see Punter (1997). Excellent
overviews stressing more the nancial engineering of such products are Doherty
(1997) and Tilley (1997). This whole area of alternative risk transfer and securitiza-
tion has become a major area of applied research in both the banking and insurance
industry. Actuaries are actively taking part in some of the new product develop-
ment, and therefore have to consider the methodological issues underlying these and
similar products.
Also recently, similar methods are being introduced in the world of nance
through the estimation of Value{at{Risk (VaR) and the so{called shortfall; see
Bassi, Embrechts and Kafetzaki (1997) and Embrechts, Samorodnitsky and Resnick
(1998). Value At Risk for End{Users (1997) contains a recent summary of some
of the more applied issues. More generally, extremes matter eminently within the
world of nance. It is no coincidence that Alan Greenspan, Chairman of the Board of
Governors of the FED, remarked at a research conference on risk measurement and
systemic risk (Washington, D.C., 16 November 1996), that \Work that characterizes
the statistical distribution of extreme events would be useful, as well."
For the general observer, extremes in the realm of nance manifest themselves
most clearly through stock market crashes or industry losses. In Figure 1, we have
 3
plotted the events leading up to and including the 1987 crash for equity data (S&P).
Extreme Value Theory (EVT) yields methods for quantifying such events and their
consequences in a statistically optimal way. See for instance McNeil (1998) for
an interesting discussion on the 1987 crash example. For a general equity book
for instance, a risk manager will be interested in estimating the resulting down{side
risk which typically can be reformulated in terms of a quantile for a Prot{and{Loss
(P & L) function.

340
320
300
S&P 500
280
260
240
220

02.01.87 02.03.87 02.05.87 02.07.87 02.09.87 02.11.87

Time

Figure 1: 1987 Crash

Another area of risk{management research where EVT is increasingly playing an
important role is in credit risk management. The interested reader may browse
J.P. Morgan's web site for information on Credit Metrics. It is no coincidence that
big investment banks are looking at actuarial methods for the sizing of reserves to
guard against future credit losses. Swiss Bank Corporation for instance introduced
ACRA (Actuarial Credit Risk Accounting) for credit risk management; see Figure 2.
In their risk measurement framework, they use the following denitions:
expected loss: the losses which must be assumed to arise on a continuing
basis as a consequence of undertaking particular business,
unexpected losses: the unusual, though predictable, losses which the Bank
should be able to absorb in the normal course of its
business,
stress loss: the possible { although improbable { extreme scenarios
which the Bank must be able to survive.
EVT oers an important set of techniques for quantifying the boundaries between
these dierent loss{classes. Moreover, EVT oers a scientic language for translating
management guidelines on these boundaries into actual numbers. A nal example
where EVT oers help is in the area of modeling default probabilities and the es-
timation of so{called diversication factors in the management of bond portfolios.
Many more examples could have been added.
 4
Frequency of loan losses

Statistical mean
Losses
500 1000 1500 2000 2500 3000

Expected loss Unexpected loss Stress loss

Absorbed by net profit Absorbed by capital Traditional

systems

Absorbed by Absorbed by ACRA reserve Absorbed by

net profit capital New Systems

Figure 2: ACRA: Actuarial Credit Risk Accounting

It is our aim in this paper to review some of the basic tools from EVT relevant
for industry wide risk management. Some examples towards the end of the paper
should give the reader a better idea of the kind of answers EVT provides. Most of the
material covered in the paper (and indeed much more) is to be found in Embrechts,
Kluppelberg and Mikosch (1997). The latter book also contains an extensive list of
further references. For reference of specic results in this book we will occasionally
refer to it as EKM () where () refers to a specic result.

2 The basic theory

As already (implicitly) stated in the introduction, the statistical analysis of extremes
is key to many of the risk management problems related to insurance, reinsurance
and nance. In order to review some of the basic ideas underlying EVT, we discuss
the most important results under the simplifying iid assumption: losses will be
assumed to be independent and identically distributed. Most of the results can be
extended to much more general models. In Example 4.2 a rst indication of such
a generalization will be given.
Throughout this paper, losses will always be denoted as positive; consequently
we concentrate in our discussion below on one{sided distribution functions (dfs) for
positive random variables (rvs).
Given basic loss data
X1 ; X2 ; : : : ; Xn iid with df F ; (1)
we are interested in the random variables
Xn;n = min (X1 ; : : : ; Xn ) ; X1;n = max (X1 ; : : : ; Xn ) : (2)
 5
Or indeed using the full set of so{called order statistics
Xn;n  Xn?1;n 


 X1;n ; (3)
we may be interested in
k
X
hr (Xr;n ) (4)
r=1

for certain functions hr , r = 1; : : : ; k and k = k(n). An important example corre-

sponds to hr  k , r = 1; : : : ; k, i.e. we average the k largest losses X ;n; : : : ; Xk;n.
1
1
Another important example would be to take k = n, hr (x) = (x ? u) where y = + +
max(0; y), for a given level u > 0. In this case we sum all excesses over u of losses
larger than u. Typically we would normalize this sum by the number of such ex-
ceedances yielding the so{called empirical mean excess function; see Example 4.1.
Most of the standard reinsurance treaties are of (or close to) the form (4). The last
example given corresponds to an excess{of{loss (XL{) treaty with priority u.
In \classical" probability theory and statistics most of the results relevant for
insurance and nance are based on sums
n
X
Sn = Xr :
r=1

Indeed the laws of large numbers, the central limit theorem (in its various degrees of
complexity), renements like Berry{Esseen, Edgeworth, saddle{point and normal{
power approximations all start from Sn{theory. Therefore, we nd in our sum{
toolkit such items like
| the normal distributions N (;  ), 2

| the {stable distributions, 0 <  < 2,

| Brownian motion,
| {stable processes, 0 <  < 2.
We are condent in our sum{toolkit when it comes to modeling/pricing/setting
reserves of random phenomena based on averages. Likewise we are condent in
statistical techniques based on these tools when applied to estimating distribution
tails \not too far" from the mean. Consider however the following easy exercise.
Exercise: It is stated that within a given portfolio, claims follow an exponential df
with mean 10 (thousand dollars, say). We have now observed 100 such claims with
largest loss 50. Do we still believe in this model? What if the largest loss would
have been 100?
 6
Solution:
The basic assumption yields that
X1 ; : : : ; X100 are iid with df P (X1  x) = 1 ? e?x=10 ; x  0 :
Therefore, for Mn = max (X ; : : : ; Xn),
1

P (M100 > x) = 1 ? (P (X1  x))100

?

= 1 ? 1 ? e?x= 10 100
:
From this, we immediately obtain
P (M100  50) = 0:4914
P (M100  100) = 0:00453 :
However, rather than doing the (easy) exact calculations above, consider the follow-
ing asymptotic argument. First, for all n  1 and x 2 R,
 
Mn
P
10
? log n  x = P (Mn  10(x + log n))
 ?x n
= 1 ? en ;

so that  
Mn
lim P ? log n  x = e?e?  (x) :x

n!1 10
Therefore, use the approximation
 x 
P (Mn  x)   ? log n
10
to obtain
P (M100  50)
 0:4902 ;
P (M  100)  0:00453 ;
100

very much in agreement with the exact calculations above. 2

Suppose we were asked the same question, but now we would have much less specic
information on F (x) = P (X  x), could we still proceed? This is exactly the point
1
where classical EVT enters. In the above exercise, we have proved the following.
Proposition 1 Suppose X ; : : : ; Xn are iid with df F  EXP(), then for x 2 R:
1

lim P (Mn ? log n  x) = (x) :

n!1
2
 7
Here are the key questions:
Q1: What is special about ? Can we get other limits, possibly for other
dfs F ?
Q2: How do we nd the norming constants  and log n in general? That
is, nd an and bn so that
 
lim P Mna? bn
n!1
x
n

exists.
Q3: Given a limit coming out of Q1, for which dfs F and norming constants
from Q2, do we have convergence to that limit? Can one say something
about second order behaviour, i.e. speed of convergence?
The solution to Q1 forms part of the famous Fisher{Tippett theorem.
Theorem 2 (EKM(Theorem 3.2.3)) Suppose X ; : : : ; Xn are iid with df F and
1
(an), (bn) are constants so that for some non{degenerate limit distribution G,
 
Mn ? b n
lim
n!1
P
an
 x = G(x) : x 2 R ;
Then G is of one of the following types:
| Type I (Frechet):
8
< 0; x0
(x) = :  > 0;
exp f?x? g ; x > 0
| Type II (Weibull):
8
< exp f?(?x) g ; x  0
(x) = :  > 0;
1; x>0

| Type III (Gumbel):


(x) = exp ?e?x ; x 2 R : 2
G is of the type H means that for some a > 0, b 2 R, G(x) = H ((x ? b)=a), x 2 R,
and the distributions of one of the above three types are called extreme value dis-
tributions. Alternatively, any extreme value distribution can be represented as
(   )
x ?  ?1=
H;; (x) = exp ? 1+  ; x 2 R:
+
 8

0.6
0.5
0.4
Gumbel

0.3
0.2
0.1 Weibull
Frechet
0.0

-5 0 5 10

Figure 3: Some examples of extreme value distributions H ;0 1; for  = 3=4 (Frechet),

 = 0 (Gumbel) and  = ?3=4 (Weibull)

Here  2 R,  2 R and  > 0. The case  > 0 ( < 0) corresponds to the Frechet
(Weibull){type df with  = 1= ( = ?1=), whereas by continuity  = 0 corre-
sponds to the Gumbel or double exponential{type df.
In Figure 3, some examples of the extreme value distributions are given. Note
that the Frechet case (the Weibull case) corresponds to a model with nite lower
(upper) bound; the Gumbel is two{sided unbounded.
Answering Q2 and Q3 is much more complicated. Below we formulate a complete
answer (due to Gnedenko) for the Frechet case. This case is the most important for
applications to (re)insurance and nance. For a general df F , we dene the inverse
of F as follows:
F (t) = inf fx 2 R : F (x)  tg ; 0 < t < 1 :
Using this notation, the p{quantile of F is dened as
xp = F (p) ; 0 < p < 1 :
Theorem 3 (EKM (Theorem 3.3.7)) Suppose X ; : : : ; Xn are iid with df F sat-
1
isfying

lim 1 ? F (tx) = x? ; x > 0 ;  > 0 : (5)

t!1 1 ? F (x)
Then for x > 0,  
Mn ? b n
lim
n!1
P
an
 x =  (x)
?

where bn = 0 and an = F 1 ? n . The converse of this result also holds true. 2
1
 9
A df F satisfying (5) is called regularly varying with index ?, denoted by F =
1 ? F 2 R? . An important consequence of the condition F 2 R? is that for
a rv X with df F ,

EX  < 1 for  <  ; (6)
= 1 for  >  :
In insurance applications, one often nds {values in the range (1; 2) whereas in
nance (equity daily log{returns say) an interval (2; 5) is common. Theorem 3 is
also reformulated as: the maximal domain of attraction of  is R? , i.e.
MDA ( ) = R? :
Dfs belonging to R? are for obvious reasons also called of Pareto type. Though we
can calculate the norming constants, the calculation of an depends on the tail of F
which in practice is unknown. The construction of MDA (  ) is also fairly easy, the
main dierence being that for F 2 MDA ( ),
xF  supfx 2 R : F (x) < 1g < 1 :
The analysis of MDA() is more involved. It contains such diverse dfs as the ex-
ponential, normal, lognormal, gamma. For details, see Embrechts, Kluppelberg and
Mikosch (1997, Section 3.3.3).

3 Tail and quantile estimation

Theorem 3 is the basis of EVT. In order to show how this theory can be put into
practice, consider for instance the pricing of an XL{treaty. Typically, the priority
(or attachment point) u is determined as a t{year event corresponding to a specic
claim event with claim size df F , say. This means that
 
u= u =F 1?
1 : (7)
t
t
In our notation used before, ut = x ? =t . Whenever t is large (typically the case in
1 1
the catastrophic (i.e. rare) event situation), the following result due to Balkema,
de Haan, Gnedenko and Pickands (see Embrechts, Kluppelberg and Mikosch (1997,
Theorem 3.4.13(b))) is very useful.
Theorem 4 Suppose X ; : : : ; Xn are iid with df F . Equivalent are:
1

i) F 2 MDA (H ),  2 R,
ii) for some function  : R ! R ,
+ +

 
lim sup Fu (x) ? G; u (x) = 0 ; (8)
u"x
( )
F 0<x<xF ?u
 10
where Fu(x) = P (X ? u  x j X > u), and the generalized Pareto df is given
by
 
x ?1=
G; (x) = 1 ? 1 +  ; (9)
+

for  > 0. 2
It is exactly the so{called excess df Fu which both risk managers as well as reinsurers
should be interested in. Theorem 4 states that for large u, Fu has a generalized
Pareto df (9). Now in order to estimate the tail F (u + x) for a xed large value of u
and all x  0 consider the trivial identity
F (u + x) = F (u) F u (x) ; u; x  0 : (10)
In order to estimate F (u + x), one rst estimates F (u) by the empirical estimator
?
Nu
F (u) b =
n
where Nu = # f1  i  n : Xi > ug. In order to have a \good" estimator for F (u),
we need u not too large: the level u has to be well within the data. Given such
a u{value, we approximate F u(x) via (8) by
?

F u (x) b = G;^ ^(u) (x)
for some estimators ^ and ^(u) depending on u. For this to work well, we need u large
(indeed in Theorem 4ii u " xF , the latter being +1 in the Frechet case). A \good"
estimator is obtained via a trade{o between these two con icting requirements
on u.
The whole statistical theory developed in order to work out the above program
runs under the name Peaks over Thresholds Method and is discussed in detail in
Embrechts, Kluppelberg and Mikosch (1997, Section 6.5), McNeil and Saladin (1997)
and references therein. Software (S{plus) implementation can be found on
http://www.math.ethz.ch/mcneil/software.
This maximum likelihood based approach also allows for modeling of the excess
intensity Nu, as well as the modeling of time (or other co{variable) dependence in
the relevant model parameters. As such, a highly versatile modeling methodology for
extremal events is available. Related approaches with application to insurance are
to be found in Beirlant, Teugels and Vynckier (1996), Reiss and Thomas (1997) and
the references therein. Interesting case studies using up{to{date EVT methodology
are McNeil (1997), Resnick (1997) and Rootzen and Tajvidi (1997). The various
steps needed to perform a quantile estimation within the above EVT context are
nicely reviewed in McNeil and Saladin (1997), where also a simulation study is to
be found. In the next section, we illustrate the methodology on real and simulated
data relevant for insurance and nance.
 11
4 Examples
4.1 Industrial re insurance data
In order to highlight the methodology brie y discussed in the previous sections, we
rst apply it to 8043 industrial re insurance claims. We show how a tail{t and the
resulting quantile estimates can be obtained. Clearly, a full analysis (as for instance
to be found in Rootzen and Tajvidi (1997) for windstorm data) would require a lot
more work. We have grouped the gures towards the end of the example.
1000
800
600
400
200
0

-5 0 5 10

Figure 4: Log{histogram of the re insurance data

Figure 4 contains the log{histogram of the data. The right{skewness stresses the
long{tailed behaviour of the underlying data. A useful plot in order to specify the
long{tailed nature of data is the so{called mean{excess plot as given in Figure 5. In
it, the mean{excess function e(u) = E (X ? u j X > u) is estimated by its empirical
counterpart n
1 X
en (u) =
# f1  i  n : X > ug (Xi ? u) :
+

i i=1
The Pareto df can be characterized by linearity (positive slope) of e(u). In general,
long{tailed dfs exhibit an upwards sloping behaviour, exponential{type dfs have
roughly a constant mean excess plot, whereas short{tailed data yield a plot decreas-
ing to 0. In our case, the upward trend clearly stresses the long{tailed behaviour.
The increase in variability towards the upper end of the plot is characteristic of the
technique, since towards the largest observation X ;n, only few data points go into
1
the calculation of en(u). The main aim of our EVT{analysis is to nd a t of the
underlying df F (x) (or of its tail F (x)) by a generalized Pareto df, especially for the
larger values of x. The empirical df F n is given in Figure 6 on a doubly logarith-
mic scale. This scale is used to highlight the tail region. Here an exact Pareto df
corresponds to a linear plot.
 12

6000
•

5000
•

Mean Excess Over Threshold

4000
•
•
• •
• •

3000
•
•
•
•
•• •

2000
•
••
•••
••••
•
••
••••••

1000
•••
•••••••
•
•••••
•
••
•••••
••••
••

0 0 500 1000 1500 2000 2500 3000

Threshold

Figure 5: Mean{excess plot of the re insurance data

1.0000

• • ••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••
••••••••••••••
••••••••••
••••••••
•••••••
•••••••
••••••
••••••
••••••
0.1000

•••••••
••••••
••••••
••••••
1 - F(x) (on log scale)

•••••
•••••
••••••
••••
••••••
0.0100

•••
••••
••••
•••••
••••
••••
•••
•••
••
••
0.0010

•• •
• •
••
•
•
•
0.0001

10^-2 10^-1 10^0 10^1 10^2 10^3 10^4

x (on log scale)

Figure 6: Empirical estimator of F on doubly logarithmic scale

Using the theory presented in Theorems 2 and 4, a maximum likelihood based

approach yields estimates for the parameters of the extreme value df H ; and the ;
generalized Pareto df G  . In order to start this procedure, a threshold value u has
;
to be chosen as estimates depend on the excesses over this threshold. The estimates
of the key shape parameter  as a function of u (alternatively, as a function of the
number of order statistics used) is given in Figure 7. Approximate 95% condence
intervals are given. The picture shows a rather stable behaviour for values of u
below 300 say. An estimate in the range (0:7; 0:9) results, which corresponds to an
{value in the range (1:1; 1:4). It should be remarked that the \optimal" value of
the threshold u to be used is dicult (if not impossible) to obtain. See Embrechts,
Kluppelberg and Mikosch (1997, p. 351) and Beirlant, Teugels and Vynckier (1996)
for some discussion. We also would like to stress that in order to produce Figure 7,
a multitude of models (one for each u chosen) has to be estimated.
For each given u, a tail{t for F u and F (as in (10)) can be obtained. For the
former, in the case of u = 100 an estimate ^ = 0:747 results. A graphical represen-
 13
695.0 224.0 141.0 88.6 68.9 51.1 42.9 36.5 30.3 27.0

3.5
3.0
Shape (xi) (CI, p = 0.95)
1.5 2.01.0
0.5 2.5

15 48 81 115 148 182 215 249 282 316 349 382 416 449 483
Exceedances

Figure 7: Maximum likelihood estimate of  as a function of the threshold u (top),

alternatively, as a function of the number of exceedances
tation of F^ is given in Figure 8. Using the parameter estimates corresponding to
100
u = 100 in (10) the tail{t of F on doubly logarithmic scale is given in Figure 9.
1.0

• • •
• • •• • •
•• •
•••
• •••
•••
0.8

• ••
• ••
•••
•
•
•••
•••
•••
0.6

••
•••
Fu(x-u)

•••
••
••••
•
•••
0.4

••
•••
•
••
••
•••
0.2

•
••
••
••••
•
••
•••
0.0

100 500 1000 5000 10000

x (on log scale)

Figure 8: Maximum likelihood t of the mean excess tail df F u based on exceedances

above u = 100
Though we have extended the generalized Pareto t to the left of u = 100, clearly
only the range above this u{value is relevant. The tting method is only designed
for the tail. Below u (where typically data are abundant) one could use a smooth
version of the empirical df. From the latter plot, quantile estimates can be deduced.
Figure 10 contains as an example the estimate for the 99.9% quantile x : to- 0 999
gether with the prole likelihood. The latter can be used to nd condence intervals
for x : . The 95% and 99% intervals are given.
0 999

Figure 11 contains the same picture but the (symmetric) condence intervals are
 14

1.0000
• • ••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••
•••••••••••••••
•••••••••
•••••••••
••••••••
•••••••
•••••••

0.1000
••••••
•••••••
••••••
••••••
•••••••
••••••

1-F(x) (on log scale)

••••••
•••••
•••••
••••••

0.0100
••••
•••
•••••
••••
••••
•••
•••
••

0.0010
•••
• • •
••
•
•
•

0.0001
•

10^-2 10^-1 10^0 10^1 10^2 10^3 10^4

x (on log scale)

Figure 9: Tail{t for F based on a threshold value of u = 100, doubly logarithmic

scale
1.0000

• • ••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••
••••••••••••••
•••••••••
••••••••
••••••••
•••••••
••••••
••••••
0.1000

••••••
•••••••
•••••
••••••
•••••••
•••••
1-F(x) (on log scale)

•••••••
••••• 95
••••••
0.0100

••••
•••••
•••••
••••
••••
••••
•••
••
•• 99
0.0010

•• • •
••
•
•
•
•
0.0001

10^-2 10^-1 10^0 10^1 10^2 10^3 10^4

x (on log scale)

Figure 10: Tail{t with an estimate for x : and the corresponding prole likelihood 0 999

calculated using the Wald statistic. Finally, the 99.9% quantile estimates across
a whole range of models (depending on the threshold value, or number of exceedances
used) are given in Figure 12. Though the estimate of x : settles between 1400 and 0 999
1500, the 95% Wald intervals are rather wide, ranging from 500 to about 2200.
The above analysis yields a summary about the high quantiles of the re insur-
ance data based on the information on extremes available in the data. The analysis
can be used as a tool in the nal pricing of risks corresponding to high layers (i.e.
catastrophic, rare events). All the methods used are based on extremes and are
fairly standard.
 15

1.0000
• • ••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••••
••••••••••••••
•••••••••
••••••••
••••••••
•••••••
••••••
••••••

0.1000
••••••
•••••••
•••••
••••••
•••••••
•••••

1-F(x) (on log scale)

•••••••
•••••
••••••

0.0100
••••
•••••
•••••
••••
••••
••••
•••
••
••

0.0010
•• • •
••
•
•
•
•

0.0001
•

10^-2 10^-1 10^0 10^1 10^2 10^3 10^4

x (on log scale)

Figure 11: Estimate of x : 0 999 with 95% Wald{statistic condence interval
695.0 224.0 141.0 88.6 68.9 51.1 42.9 36.5 30.3 27.0
2500 2000
0.999 Quantile (CI, p = 0.95)
1000 1500 500

15 48 81 115 148 182 215 249 282 316 349 382 416 449 483
Exceedances

Figure 12: Estimates of the quantile x : as a function of the threshold u. The 0 999
vertical line indicates the model corresponding to u = 100

4.2 An ARCH{example
To illustrate further some of the available techniques, we simulated an ARCH(1)
time series of length 99,000. The time series, called testarch, has the form
?
1=2
n = Xn  + n2?1 ; n  1; (11)
where fXng are iid N (0; 1) random variables. In our simulation, we took
 = 1 ;  = 0:5 :
From known results of Kesten (1973) (see also Embrechts, Kluppelberg and Mikosch
(1997, Theorem 8.4.12), Goldie (1991), Vervaat (1979))
P (1 > x)  cx?2 ; x ! 1 ; (12)
 16
and we get from Table 3.2 of de Haan et al. (1989) that
 = 2:365
(see also Hooghiemstra and Meester (1995)).
There are several reasons why we choose to simulate an ARCH process.
| Despite the fact that the ARCH process is dependent, much of the classical
extreme value analysis applies with suitable modications.
| The ARCH process has heavy tails which matches what is observed in data
sets emerging from nance.
| Although it is often plausible to model large insurance claims as iid, data
from the nance industry such as exchange rate data are demonstrably not
iid. Some of these examples have the property that the data look remarkably
uncorrelated but squares or absolute values of the data appear to have high
correlations. It is this property that the ARCH process and its cousins were
designed to model. See for instance Taylor (1986) for more details.
To experiment with this ARCH data we took the rst 10,000 observations to
form a data set shortarch which will be used for estimation. Then based on the
estimation, some model based predictions can be made and compared with actual
data in testarchnshortarch.
Figure 13 shows a time series plot of shortarch. The plot exhibits the character-
istic heavy tail appearance.
shortarch
10
5
0
-5
-10

0 2000 4000 6000 8000 10000

Figure 13: Time series plot of shortarch

The Hill estimator is a popular way of detecting heavy tails and estimating
the Pareto index 1= for  > 0 in (9). See for instance Embrechts, Kluppelberg
and Mikosch (1997) for an introduction. Figure 14 contains four views of the Hill
estimator applied to shortarch.
 17
Hill plot AltHill

2 3 4 5 6 7

2 3 4 5 6 7
Hill estimate of alpha

Hill estimate of alpha

0 500 1000 1500 2000 0.2 0.4 0.6 0.8
number of order statistics theta

AltsmooHill AltHill and AltsmooHill

2 3 4 5 6 7
6
Hill estimate of alpha

Hill estimate of alpha

3 4 5

0.2 0.4 0.6 0.2 0.4 0.6 0.8

theta theta

Figure 14: Hill plots of shortarch

The Hill estimator is known to be a consistent estimator of 2 for the ARCH

process (Resnick and Starica (1996)). In our case, the Hill estimator is trying to
estimate 2  22:365  4:7. A review of Figure 14 yields an estimate of about 4. If
Hk;n represents the Hill estimator when the sample is n and k upper order statistics
are used, i.e. !?
k 1

H =
1 X
log X ? log X ;
k;n j;n k;n
k j =1
? ; 1  k  n . The upper ?

the usual methodology is to make a Hill plot k; Hk;n 1

left graph is a Hill plot with some values for small and large k deleted to make
the picture scale attractively. The upper 
right plot isthe Hill plotin alt scale (see
Resnick and Starica, 1997) where we plot ; H[?n ];n ; 0    1 . The lower left 1

plot applies a smoother (Resnick and Starica, 1997) and plots in alt scale.
A supplementary tool for estimating the Pareto index is the QQ plot (see Em-
brechts, Kluppelberg and Mikosch (1997, Section 6.2.1)) and this has the added
advantage that it allows simultaneous estimation of the constant c appearing in
(12). The method is sensitive to the choice of the number of upper order statistics
and some trial and error is usually necessary. In Figure 15 we give the QQ plot
based on the upper 400 order statistics. This gives estimates of 2 = 3:851904 and
c = 1:15389. (Applying this technique to the full testarch data produced estimates
of 2 = 3:861008 and c = 1:319316, when the number of upper order statistics was
 18
300.)

•
• •

2.0
•
• • •
•
•• • •

log sorted data

•
•
••••
••
•••
•••
••••••

1.5
••••
•••
••••••••
••
•••••
••••
•••••••
•••
•••••••••
•
•••••
••••••
•••••
•••••
•
••
•••••••••
••••
••••••
1.0
••••••
•••••
••
••
••••••••
••

3 4 5 6 7 8 9
quantiles of exponential

Figure 15: QQ plot of shortarch

Based on these estimates, we experiment with some predictions and compare
these predictions with what is observed from the part of the data set testarch, called
playarch obtained by removing the 10,000 shortarch observations. Thus the length
of playarch is 99; 000 ? 10; 000 = 89; 000. In the following table we give estimated
marginal probabilities that the ARCH variable exceeds x for x = 5; 10; 15; 20. Note
that we are predicting values that are beyond the range of the data and have not
been observed. The second row gives the estimate (12) based on the tted values
for c and 2. In the third row we compute the empirical frequency that elements of
playarch exceed x. The last row gives the corresponding probabilities 1 ?  (x; ;  ) 2

based on a normal distribution whose mean and variance are the sample mean
and variance computed from shortarch. One can see from the table the penalty
paid for ignoring extreme value analysis and relying on more conventional normal
distribution based analysis.
x 5 10 15 20
P (X > x) 0.002309062 0.0001589119 0.0000332 0.0000109
P (X > x)
b 0.002640449 0.0001797753 0.0000449 0.0000112
1 ?  (x; ;  ) 0.000186 5:29  10?
2
0 0 13

The extreme value theory for the ARCH process is somewhat complicated by
the ARCH dependence structure not present for an iid sequence. A quantity called
the extremal index must be accounted for; see Embrechts, Kluppelberg and Mikosch
(1997, Section 8.1). From (11) and de Haan et al. (1989), Table 3.2 we have
 
1
P (max fi ; : : : ; ng  y )  exp ? c0 ny ?  ; 2
(13)
2
where the extremal index 0 = 0:835 accounts for the eective reduction in sample
size due to dependence. From this formula, estimates of upper quantiles can be
 19
worked out. The upper 100p% quantile xp would be
 ?1=2
xp  ? log(1 ? p) 20c n
: (14)
We give a few representative values:
y 15 20 25 30
P (max fi ; : : : ; ng  y ) 0.28229 0.65862 0.83794 0.91613

p 0.05 0.01 0.005 0.0005

xp 34.47153 52.63015 63.04769 114.692

4.3 Value{at{Risk: a word of warning

We already pointed out the similarity in estimating attachment points or retentions
in reinsurance and VaR calculations in nance. Both are statistically based methods,
where the basic underlying risk measure corresponds to a quantile estimate xbp of
an unknown df. Through the work of Artzner et al. (1996, 1998) we know that
a quantile based risk measure for general (non{normal) data fails to be coherent,
i.e. such a measure is not subadditive creating inconsistencies in the construction of
risk capital based upon them. This situation typically occurs in portfolios containing
non{linear derivatives. Further critical statements concerning VaR are to be found
in Danielsson, Hartmann and de Vries (1992), Garman (1997), Longin (1997a,b)
and Cardenas et al. (1997).
A much better, and indeed (almost) coherent risk measure is the conditional
VaR (or so{called mean excess)
E (X j X > xbp ) : (15)
For the precise formulation of the \almost" bove, see Artzner et al. (1996). We
only want to point out that the latter measure is well{known in insurance, but
only gradually is being recognized as fundamental to risk management. It is one
of the (many) examples where an exchange of ideas between actuaries and nance
experts may lead to improved risk measurement. Note that in the equation above,
E (X j X > xbp ) = e (xbp )+xbp . One could use the mean excess plot f(u; en (u)) ; u  0g
in order to visualize the tail behavior of the underlying data and hence get some
insight on the coherent risk measure (15) above.
As a nal example we have plotted in Figure 16 the daily log{returns of BMW
over the period January 23, 1973|July 12, 1996, together with the mean excess plot
of the absolute values of the negative returns (hence corresponding to downside risk).
In the latter plot, the heavy{tailed nature of the returns is very clear. Also clear
is the change in curvature of the plot around 0.03. This phenomenon is regularly
observed in all kinds of data. One can look at it as a pictorial view of Theorem 4:
indeed, Smith (1990) indicates how to base on this observation a graphical tool
to determine an initial threshold value for an extreme value analysis. See also
 20

0.05
-0.15 -0.05
02.01.73 02.01.77 02.01.81 02.01.85 02.01.89 02.01.93
Time

0.030
•
•
• •
Mean Excess

•••• •
•
• •
••••• •••
0.020
• ••
•••••• ••
• •••
• ••
•••••••••••••••••••
• ••
• •
••• •
•
••••••••••••• ••
••••••••••• •
0.010

••••••••••••••••••••••••••••••••••••••••••••••••••••••••••
0.0 0.02 0.04 0.06 0.08 0.10
Threshold

Figure 16: Time series and mean excess plots of BMW return data

Embrechts, Kluppelberg and Mikosch (1997, p. 356) for a discussion. We would like
to warn the reader however that due to the intricate dependencies in nance data,
one should be careful in using these plots beyond the mere descriptive level.
More work is needed to combine the ideas presented in this paper with detailed
statistical information on nancial time series before risk measures such as condi-
tional VaR (15) can be formulated precisely and estimated reliably. Once more,
the interplay between statisticians, nance experts and actuaries should prove to be
fruitful towards achieving this goal.

References
Artzner, P., Delbaen, F., Eber, J.-M. and D. Heath (1996) A characterization of
measures of risk. Universite de Strasbourg preprint.
Artzner, P., Delbaen, F., Eber, J.-M. and D. Heath (1998) Thinking coherently.
RISK, 10, 68{71.
Bassi, F., Embrechts, P. and M. Kafetzaki (1997) Risk Management and quantile
estimation. In: Adler, R., Feldman, R. and Taqqu, M.S. (Eds.) A Practi-
cal Guide to Heavy Tails: Statistical Techniques for Analysing Heavy{Tailed
Distributions. Birkhauser, Boston.
Beirlant, J., Teugels, J.L.T. and P. Vynckier (1996) Practical analysis of extreme
values. Leuven University Press, Leuven.
Canter, M.S., Cole, J.B. and R.L. Sandor (1996) Insurance derivatives: a new asset
class for the capital markets and a new hedging tool for the insurance industry.
Journal of Derivatives, 89{104.
 21
Cardenas, J., Fruchard, E., Koehler, E., Michel, C. and I. Thomazeau (1997) VaR:
One step beyond. RISK 10, 72{74.
Danielsson, J., Hartmann, P. and C.G. de Vries (1998) The cost of conservatism.
RISK 11, 101{103.
Doherty, N.A. (1997) Financial innovation in the management of catastrophe risk.
Joint Day Proceedings, XXVIIIth International ASTIN Colloquium and 7th
International AFIR Colloquium, Cairns (Australia), 1{26.
Embrechts, P., Kluppelberg, C. and T. Mikosch (1997) Modelling extremal events
for insurance and nance. Springer{Verlag, Berlin.
Embrechts, P., Samorodnitsky, G. and S.I. Resnick (1998) Living on the edge.
RISK 11, 96{100.
Garman, M. (1997) Taking VaR to pieces. RISK 10.
Goldie, C.M. (1991) Implicit renewal theory and tails of solutions of random equa-
tions. Annals of Applied Probability 1, 126{166.
Haan, L. de, Resnick, S.I., Rootzen, H. and C.G. de Vries (1989) Extremal be-
haviour of solutions to a stochastic dierence equation with applications to
ARCH processes. Stoch. Proc. Appl. 32, 213{224.
Hooghiemstra, G. and L. Meester (1995) Computing the extremal index of special
Markov chains and queues. Stochastic Process. Appl. 65, 171{185.
Jae, D.M. and T. Russell (1996) Catastrophe insurance, capital markets and
uninsurable risk. Financial Institutions Center, The Wharton School, 96{12.
Kesten, H. (1973) Random dierence equations and renewal theory for products of
random matrices. Acta Math. 131, 207{248.
Longin, F.M. (1997a) From value at risk to stress testing: the extreme value ap-
proach. ESSEC and CEPR Paris preprint.
Longin, F.M. (1997b) Beyond the VaR. ESSEC and CEPR Paris preprint.
McNeil, A.J. (1997) Estimating the tails of loss severity distributions using extreme
value theory. ASTIN Bulletin 27(2), 117{137.
McNeil, A.J. (1998) On extremes and crashes. RISK 11, 99.
McNeil, A.J. and T. Saladin (1997) The peaks over threshold method for estimating
high quantiles of loss distributions. Proceedings of the XXVIIIth International
ASTIN Colloquium, Cairns, 23{43.
Punter, A. (1997) A critical evaluation of progress to date. In: Practical appli-
cations of nancial market tools to corporate risk management. International
Risk Management, p. 7{12.
 22
Reiss, R.{D. and M. Thomas (1997) Statistical Analysis of Extremal Values, Birk-
hauser, Basel.
Resnick, S.I. (1997) Discussion of the Danish data on large re insurance losses.
ASTIN Bulletin 27(2), 139{151.
Resnick, S.I. and C. Starica (1996) Tail index estimation for dependent data. Avail-
able as TR1174.ps.Z at http://www.orie.cornell.edu/trlist/trlist.html.
To appear in Ann. Applied Probability.
Resnick, S.I. and C. Starica (1997) Smoothing the Hill estimator. Adv. Applied
Probability 29, 271{293.
Rootzen, H. and N. Tajvidi (1997) Extreme value statistics and windstorm losses:
a case study. Scand. Actuarial J. 1, 70{94.
Schmock, U. (1997) Estimating the value of the Wincat coupons of the Winterthur
Insurance convertible bond. Joint Day Proceedings, XXVIIIth International
ASTIN Colloquium and 7th International AFIR Colloquium, Cairns (Aus-
tralia), 231{259.
Sigma (1995) Natural catastrophes and major losses in 1994: Third highest loss
burden in the history of insurance. Swiss Reinsurance Company, 3/95.
Sigma (1997) Natural catastrophes and major losses in 1996: high losses from man{
made disasters, but not extremely costly losses from natural catastrophes.
Swiss Reinsurance Company, 3/97.
Smith, R.L. (1990) Extreme value theory. In: Ledermann, W. (Chief Ed.) Hand-
book of Applicable Mathematics, Supplement, pp. 437{472. Wiley, Chichester.
Taylor, S.J. (1986) Modelling Financial Time Series, Wiley, Chichester.
Tilley, J.A. (1997) The securitization of catastrophic property risks. Joint Day Pro-
ceedings, XXVIIIth International ASTIN Colloquium and 7th International
AFIR Colloquium, Cairns (Australia), 27{53.
Value at Risk for End{Users (1997) Risk Magazine, Supplement, March 1997.
Paul Embrechts Sidney I. Resnick and Gennady Samorodnitsky
Department of Mathematics School of Operations Research
ETHZ and Industrial Engineering
CH { 8092 Zurich, Switzerland Cornell University
e-mail: embrechts@math.ethz.ch Rhodes Hall/ETC Building
Ithaca, New York 14853, USA
e-mail: sid@orie.cornell.edu
e-mail: gennady@orie.cornell.edu