Professional Documents
Culture Documents
Embrechts ExtremeValue
Embrechts ExtremeValue
Y
Copyright 2010 by the Society of Actuaries, Schaumburg, Illinois.
Posted with permission.
Name /8042/03 04/21/99 09:19AM Plate # 0 NAAJ (SOA) pg 30 # 1
ABSTRACT
The financial industry, including banking and insurance, is undergoing major changes. The
(re)insurance industry is increasingly exposed to catastrophic losses for which the requested cover
is only just available. An increasing complexity of financial instruments calls for sophisticated risk
management tools. The securitization of risk and alternative risk transfer highlight the conver-
gence of finance and insurance at the product level. Extreme value theory plays an important
methodological role within risk management for insurance, reinsurance, and finance.
1. INTRODUCTION Table 1
Consider the time series in Table 1 of loss ratios California Earthquake Data
(yearly data) for earthquake insurance in California
from 1971 through 1993. The data are taken from 1971 17.4 1979 2.2 1987 22.8
1972 0.0 1980 9.2 1988 11.5
Jaffe and Russell (1996). 1973 0.6 1981 0.9 1989 129.8
On the basis of these data, who would have guessed 1974 3.4 1982 0.0 1990 47.0
the 1994 value of 2272.7? Indeed, on the 17th of Jan- 1975 0.0 1983 2.9 1991 17.2
1976 0.0 1984 5.0 1992 12.8
uary of that year the 6.6-Richter-scale Northridge 1977 0.7 1985 1.3 1993 3.2
earthquake hit California, causing an insured damage 1978 1.5 1986 9.3
of $10.4 billion and a total damage of $30 billion,
making 1994 the year with the third highest loss bur-
den (natural catastrophes and major losses) in the his- man-made catastrophes. For the United States alone,
tory of insurance. The front-runners are 1992 (the Canter, Cole, and Sandor (1996) estimate an approx-
year of hurricane Andrew) and 1990 (the year of the imate $245 billion of capital in the insurance and re-
winter storms Daria and Vivian). For details on these, insurance industry to service a country that has $25
see Sigma (1995, 1997). 30 trillion worth of property. It is no surprise that the
The reinsurance industry experienced a rise in both finance industry has seized upon this by offering (of-
intensity and magnitude of losses due to natural and ten in joint ventures with the (re)insurance world)
properly securitized products in the realm of catastro-
phe insurance. New products are being born at an in-
creasing pace. Some of them have only a short life,
* A first version of this paper was presented by the first author as an
invited paper at the XXVIIIth International ASTIN Colloquium in others are reborn under a different shape, and some
Cairns and published in the conference proceedings under the title do not survive. Examples include:
Extremes and Insurance. Catastrophe (CAT) futures and PCS options (Chi-
Paul Embrechts, Ph.D., is Professor in the Department of Mathe- cago Board of Trade). In these cases, securitization
matics, ETHZ, CH-8092 Zurich, Switzerland, e-mail, embrechts@ is achieved through the construction of derivatives
math.ethz.ch. written on a newly constructed industry-wide loss-
Sidney I. Resnick, Ph.D., is Professor in the School of Operations
ratio index.
Research and Industrial Engineering, Cornell University, Rhodes Hall /
ETC Building, Ithaca, New York 14853, e-mail, sid@orie.cornell. Convertible CAT bonds. The Winterthur convertible
edu. hail-bond is an example. This European-type con-
Gennady Samorodnitsky, Ph.D., is Associate Professor in the School vertible has an extra coupon payment contingent on
of Operations Research and Industrial Engineering, Cornell Univer- the occurrence of a well-defined catastrophic (CAT)
sity, Rhodes Hall / ETC Building, Ithaca, New York 14853, e-mail, event: an excessive number of cars in Winterthurs
gennady@orie.cornell.edu.
30
Name /8042/03 04/21/99 09:19AM Plate # 0 NAAJ (SOA) pg 31 # 2
Swiss portfolio damaged in a hail storm over a spe- theory (EVT) yields methods for quantifying such
cific time period. For details, see Schmock (1997). events and their consequences in a statistically opti-
Further interesting new products are the multiline, mal way. (See McNeil 1998 for an interesting discus-
multiyear, high-layer (infrequent event) products, sion of the 1987 crash example.) For a general equity
credit lines, and the catastrophe risk exchange (CA- book, for instance, a risk manager will be interested
TEX). For a brief review of some of these instruments, in estimating the resulting down-side risk, which typ-
see Punter (1997). Excellent overviews stressing the ically can be reformulated in terms of a quantile for a
financial engineering of such products are Doherty profit-and-loss function.
(1997) and Tilley (1997). Alternative risk transfer and EVT is also playing an increasingly important role
securitization have become major areas of applied re- in credit risk management. The interested reader
search in both the banking and insurance industries. may browse J.P. Morgans web site (http://www.
Actuaries are actively taking part in some of the new jpmorgan.com) for information on CreditMetrics. It is
product development and therefore have to consider no coincidence that big investment banks are looking
the methodological issues underlying these and simi- at actuarial methods for the sizing of reserves to guard
lar products. against future credit losses. Swiss Bank Corporation,
Also, similar methods have recently been intro- for instance, introduced actuarial credit risk account-
duced into the world of finance through the estima- ing (ACRA) for credit risk management; see Figure 2.
tion of value at risk (VaR) and the so-called shortfall; In their risk measurement framework, they use the
see Bassi, Embrechts, and Kafetzaki (1998) and following definitions:
Embrechts, Samorodnitsky, and Resnick (1998). Expected loss: the losses that must be assumed to
Value At Risk for End-Users (1997) contains a re- arise on a continuing basis as a consequence of un-
cent summary of some of the more applied issues. dertaking particular business
More generally, extremes matter eminently within the Unexpected loss: the unusual, though predictable,
world of finance. It is no coincidence that Alan Green- losses that the bank should be able to absorb in the
span, chairman of the U.S. Federal Reserve, remarked normal course of its business
at a research conference on risk measurement and Stress loss: the possiblealthough improbable
systemic risk (Washington, D.C., November 1995) extreme scenarios that the bank must be able to
that Work that characterizes the statistical distri- survive.
bution of extreme events would be useful, as well. EVT offers an important set of techniques for quan-
For the general observer, extremes in the realm of tifying the boundaries between these different loss
finance manifest themselves most clearly through classes. Moreover, EVT provides a scientific language
stock market crashes or industry losses. In Figure 1, for translating management guidelines on these
we have plotted the events leading up to and including
the 1987 crash for equity data (S&P). Extreme value
Figure 2
Actuarial Credit Risk Accounting (ACRA)
Figure 1
1987 Crash
Name /8042/03 04/21/99 09:19AM Plate # 0 NAAJ (SOA) pg 32 # 3
boundaries into actual numbers. Finally, EVT helps in excesses over u of losses larger than u. Typically we
the modeling of default probabilities and the estima- would normalize this sum by the number of such ex-
tion of diversification factors in the management of ceedances yielding the so-called empirical mean ex-
bond portfolios. Many more examples can be added. cess function; see Section 4.1. Most of the standard
It is our aim in this paper to review some of the reinsurance treaties are of (or close to) the form (4).
basic tools from EVT relevant for industry-wide inte- The last example given corresponds to an excess-of-
grated risk management. Some examples toward the loss (XL) treaty with priority u.
end of this paper will give the reader a better idea of In classical probability theory and statistics most
the kind of answers EVT provides. Most of the material of the results relevant for insurance and finance are
covered here (and indeed much more) is found in Em- based on sums
OX.
brechts, Kluppelberg, and Mikosch (1997), which also n
contains an extensive list of further references. For Sn 5 r
reference to a specific result in this book, we will oc- r51
casionally identify it as EKM. Indeed, the laws of large numbers, the central limit
theorem (in its various degrees of complexity), refine-
2. THE BASIC THEORY ments like Berry-Esseen, Edgeworth, and saddle-point,
and normal-power approximations all start from Sn
The statistical analysis of extremes is key to many of
theory. Therefore, we find in our toolkit for sums such
the risk management problems related to insurance,
items as the normal distributions N(m, s 2); the a-
reinsurance, and finance. In order to review some of
stable distributions, 0 , a , 2; Brownian motion; and
the basic ideas underlying EVT, we discuss the most
a-stable processes, 0 , a , 2.
important results under the simplifying iid assump-
We are confident of our toolkit for sums when it
tion: losses will be assumed to be independent and
comes to modeling, pricing, and setting reserves of
identically distributed. Most of the results can be ex-
random phenomena based on averages. Likewise we
tended to much more general models. In Section 4.2
are confident of statistical techniques based on these
a first indication of such a generalization will be given.
tools when applied to estimating distribution tails
Throughout this paper, losses will always be denoted
not too far from the mean. Consider, however, the
as positive; consequently we concentrate in our dis-
following easy exercise.
cussion below on one-sided distribution functions
(dfs) for positive random variables (rvs). Exercise
Given basic loss data
It is stated that, within a given portfolio, claims follow
X1, X2, . . . , Xn iid with df F, (1) an exponential df with mean 10 (thousand dollars,
say). We have now observed 100 such claims with larg-
we are interested in the random variables
est loss 50. Do we still believe in this model? What if
Xn,n 5 min(X1, . . . . , Xn), X1,n 5 max(X1, . . . , Xn). the largest loss would have been 100?
(2) Solution
Or, indeed, using the full set of so-called order The basic assumption yields that
statistics
X1, . . . , X100 are iid with df P(X1 # x)
Xn,n # Xn21,n # z z z # X1,n , (3)
5 1 2 e 2x/10, x $ 0.
we may be interested in
Therefore, for Mn 5 max(X1, . . . , Xn),
O
k
hr(Xr, n) (4) P(M100 . x) 5 1 2 (P(X1 # x))100
r51
However, rather than doing the (easy) exact calcula- Theorem 2 (EKM Theorem 3.2.3)
tions above, consider the following asymptotic argu- Suppose X1, . . . , Xn are iid with df F and (an), (bn)
ment. First, for all n $ 1 and x [ R, are constants so that for some nondegenerate limit
S
Mn
D distribution G,
S D
P 2 log n # x 5 P(Mn # 10(x 1 log n))
10 Mn 2 bn
#x 5 G(x), x [ R.
S D
lim P
e 2x n
n` an
5 12 ,
n Then G is of one of the following types:
Type I (Frechet):
so that
lim P
n`
S Mn
10
2 log n # x D 5 e 2e
2x
[ L(x).
Fa(x) 5 H
0, x#0
exp{2x 2a}, x . 0
a.0
Type II (Weibull):
H
Therefore, use the approximation
S D
exp{2(2x)a}, x # 0
x Ca(x) 5 a.0
P(Mn # x) < L 2 log n 1, x.0
10
Type III (Gumbel):
to obtain
L(x) 5 exp{2e 2x}, x [ R. M
P(M100 $ 50) < 0.4902,
P(M100 $ 100) < 0.00453, G is of the type H means that for some a . 0, b [ R,
G(x) 5 H((x 2 b)/a), x [ R, and the distributions
very much in agreement with the exact calculations of one of the above three types are called extreme
above. value distributions. Alternatively, any extreme value
Suppose we were asked the same question but had distribution can be represented as
HS D J
much less specific information on F(x) 5 P (X1 # x); 21/j
x2m
could we still proceed? This is exactly the point where Hj ; m,s (x) 5 exp 2 11j , x [ R.
classical EVT enters. In the above exercise, we have s 1
lim P
n`
SMn 2 bn
an
#x D
exists?
Q3: Given a limit coming out of Q1, for which dfs F
and norming constants from Q2, do we have con-
vergence to that limit? Can one say something
about second order behavior, that is, speed of
convergence?
The solution to Q1 forms part of the famous Gne-
denko, Fisher-Tippett theorem.
Name /8042/03 04/21/99 09:19AM Plate # 0 NAAJ (SOA) pg 34 # 5
corresponds to the Gumbel, or double exponential- The analysis of MDA(L) is more involved. It contains
type, df. such diverse dfs as the exponential, normal, lognor-
In Figure 3, some examples of the extreme value mal, and gamma. For details see Embrechts, Kluppel-
distributions are given. Note that the Frechet case berg, and Mikosch (1997, Section 3.3.3).
(the Weibull case) corresponds to a model with finite
lower (upper) bound; the Gumbel case is two-sided
unbounded. 3. TAIL AND QUANTILE ESTIMATION
Answering Q2 and Q3 is much more complicated. Theorem 3 is the basis of EVT. In order to show how
Below we formulate a complete answer (due to Gne- this theory can be put into practice, consider, for in-
denko) for the Frechet case. This case is the most stance, the pricing of an XL treaty. Typically, the pri-
important for applications to (re)insurance and fi- ority (or attachment point) u is determined as a t-year
nance. For a general df F, we define the inverse of F event corresponding to a specific claim event with
as: claim size df F, for example. This means that
lim P
n`
SMn 2 bn
an
#x D 5 Fa(x),
ii) for some function b : R1 R1,
lim sup u Fu(x) 2 G j, b(u)(x)u 5 0, (8)
uxF 0,x,xF2u
where bn 5 0 and an 5 F (1 2 1/n). The converse
of this result also holds true. M where Fu(x) 5 P(X 2 u # x u X . u), and the gen-
eralized Pareto df is given by
A df F satisfying (5) is called regularly varying with
index 2a, denoted by F 5 1 2 F [ R2a. An important
consequence of the condition F [ R2a is that for a rv
Gj , b(x) 5 1 2 S11j
x
bD 21/j
1
, (9)
X with df F, for b . 0. M
EXb H, ` for b , a,
5 ` for b . a.
(6) It is exactly the so-called excess df Fu that risk man-
agers as well as reinsurers should be interested in.
In insurance applications, one often finds a-values in Theorem 4 states that for large u, Fu has a generalized
the range (1, 2), whereas in finance (equity daily log- Pareto df (9). Now, to estimate the tail F (u 1 x) for
returns, say) an interval (2, 5) is common. Theorem a fixed large value of u and all x $ 0, consider the
3 is also reformulated thus: The maximal domain of trivial identity
attraction of Fa is R2a, that is,
F(u 1 x) 5 F (u) Fu(x), u, x $ 0. (10)
MDA(Fa) 5 R2a.
In order to estimate F (u 1 x), one first estimates
Dfs belonging to R2a are for obvious reasons also F (u) by the empirical estimator
called Pareto type. Though we can calculate the norm-
N
ing constants, the calculation of an depends on the (F(u)) 5 u,
tail of F, which in practice is unknown. The construc- n
tion of MDA (Ca) is also fairly easy, the main differ- where Nu 5 # {1 # i # n : Xi . u}. In order to have
ence being that for F [ MDA(Ca), a good estimator for F (u), we need u not too large:
xF [ sup{x [ R : F (x) , 1} , `.
Name /8042/03 04/21/99 09:19AM Plate # 0 NAAJ (SOA) pg 35 # 6
the level u has to be well within the data. Given such Figure 4
a u-value, we approximate Fu(x) via (8) by Log Histogram of the Fire Insurance Data
(Fu(x)) 5 G j , b(u)
(x)
for some estimators j and b (u), depending on u. For
this to work well, we need u large (indeed, in Theorem
(4ii), u x F , the latter being 1` in the Frechet case).
A good estimator is obtained via a trade-off between
these two conflicting requirements on u.
The statistical theory developed to work out the
above program runs under the name Peaks over
Thresholds Method and is discussed in detail in Em-
brechts, Kluppelberg, and Mikosch (1997, Section
6.5), McNeil and Saladin (1997), and references
therein. Software (S-plus) implementation can be
found at
http://www.math.ethz.ch/,mcneil/software.
e(u) 5 E(X 2 u u X . u) is estimated by its empirical
This maximum-likelihood-based approach also allows
counterpart
for modeling of the excess intensity Nu , as well as the
O (X 2 u) .
n
modeling of time (or other co-variable) dependence in 1 1
the relevant model parameters. As such, a highly ver- e n(u) 5
#{1 # i # n : X i . u} i51
i
4. EXAMPLES
Figure 5
4.1 Industrial Fire Insurance Data Mean-Excess Plot of the Fire Insurance Data
In order to highlight the methodology briefly dis-
cussed in the previous sections, we first apply it to
8043 industrial fire insurance claims. We show how a
tail-fit and the resulting quantile estimates can be ob-
tained. Clearly, a full analysis (as found, for instance,
in Rootzen and Tajvidi 1997 for windstorm data)
would require much more work.
Figure 4 contains the log histogram of the data. The
right-skewness stresses the long-tailed behavior of
the underlying data. A useful plot for specifying the
long-tailed nature of data is the mean-excess plot
given in Figure 5. In it, the mean-excess function
Name /8042/03 04/21/99 09:19AM Plate # 0 NAAJ (SOA) pg 36 # 7
Figure 11
Figure 8
Estimate of x 0.999 with 95% Wald-Statistic
Maximum Likelihood Fit of the Mean-Excess
Confidence Interval
Tail Fu Based on Exceedances above u 5 100
Figure 12
Figure 9 Estimates of the Quantile x 0.999 as a Function
Tail Fit for F Based on a Threshold Value of of the Threshold u
u 5 100, Doubly Logarithmic Scale
j51
j ,n D
2 log Xk,n
21
,
applies with suitable modifications.
the usual methodology is to make a Hill plot {(k,
The ARCH process has heavy tails, which matches
what is observed in data sets emerging from finance. H 21
k,n ), 1 # k # n}. The upper left graph is a Hill plot
with some values for small and large k deleted to make
Although it is often plausible to model large insur-
ance claims as iid, data from the finance industry the picture scale attractively. The upper right plot is
such as exchange rate or equity data are demon- the Hill plot in alt scale (see Resnick and Starica
strably not iid. For some of these examples, the data 1997), where we plot {(u, H 21 [nu],n), 0 # u # 1}. The
look remarkably uncorrelated, but squares or abso- lower left plot applies a smoother (Resnick and Star-
lute values of the data appear to have high correla- ica 1997) and plots in alt scale.
tions. It is this property that the ARCH process and A supplementary tool for estimating the Pareto in-
its cousins were designed to model. See, for in- dex is the QQ plot (see Embrechts, Kluppelberg, and
stance, Taylor (1986) for more details. Mikosch 1997, Section 6.2.1), which has the added
To experiment with these ARCH data we took the advantage of allowing simultaneous estimation of the
first 10,000 observations to form a data set short- constant c appearing in (12). The method is sensitive
arch, which will be used for estimation. Based on the to the choice of the number of upper order statistics,
estimation, some model-based predictions can then and some trial and error is usually necessary. In Figure
be made and compared with actual data in 15 we give the QQ plot based on the upper 400 order
testarch\shortarch. statistics. This gives estimates of 2k 5 3.851904 and
Figure 13 shows a time series plot of shortarch. The c 5 1.15389. (Applying this technique to the full test-
plot exhibits the characteristic heavy tail appearance. arch data produced estimates of 2k 5 3.861008 and
The Hill estimator is a popular way of detecting c 5 1.319316, when the number of upper order sta-
heavy tails and estimating the Pareto index 1/j for j . tistics was 300.)
0 in (9). See, for instance, Embrechts, Kluppelberg, Based on these estimates, we experiment with some
and Mikosch (1997) for an introduction. Figure 14 predictions and compare them with what is observed
contains four views of the Hill estimator applied to from that part of the data set testarch called playarch,
shortarch. obtained by removing the 10,000 shortarch observa-
The Hill estimator is known to be a consistent es- tions. Thus the length of playarch is 99,000 2
timator of 2k for the ARCH process (Resnick and Star- 10,000 5 89,000. In Table 2 we give estimated mar-
ica 1996). In our case, the Hill estimator is trying to ginal probabilities that the ARCH variable exceeds x
estimate 2k < 2 3 2.365 < 4.7. A review of Figure for x 5 5, 10, 15, 20. Note that we are predicting
14 yields an estimate of about 4. If Hk,n represents the values that are beyond the range of the data and have
Hill estimator when the sample is n and k upper order not been observed. The second row gives the estimate
statistics are used, that is, (12) based on the fitted values for c and 2k. In the
third row we compute the empirical frequency that
elements of playarch exceed x. The last row gives the
Figure 13 corresponding probabilities 1 2 F(x, m, s 2) based on
Time Series Plot of Shortarch a normal distribution whose mean and variance are
the sample mean and variance computed from short-
arch. One can see from Table 2 the penalty paid for
ignoring extreme value analysis and relying on more
conventional normal-distribution-based analysis.
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
Name /8042/03 04/21/99 09:19AM Plate # 0 NAAJ (SOA) pg 39 # 10
Figure 14
Hill Plots of Shortarch
P(max{j i, . . . , j n} # y)
xp < S 2log(1 2 p)
2
D 21/2k
H J
. (14)
cu9n
1
< exp 2 cu9ny 22k , (13)
2 Table 3 gives a few representative values.
where the extremal index u9 5 0.835 accounts for the
effective reduction in sample size due to dependence. 4.3 Value at Risk: A Word of Warning
From this formula, estimates of upper quantiles can We have already pointed out the similarity in estimat-
be worked out. The upper 100p% quantile xp would be ing attachment points or retentions in reinsurance
and VaR calculations in finance. Both are statistically
based methods, where the basic underlying risk mea-
sure corresponds to a quantile estimate x p of an un-
Figure 15 known df. Through the work of Artzner et al. (1996,
QQ Plot of Shortarch 1998) we know that a quantile-based risk measure for
general (nonnormal) data fails to be coherentsuch
a measure is not subadditive, creating inconsistencies
in the construction of risk capital based upon it. This
Table 2
Exceedance Probabilities
for the ARCH Example
x 5 10 15 20
P(X . x) 0.002309062 0.0001589119 0.0000332 0.0000109
P (X . x) 0.002640449 0.0001797753 0.0000449 0.0000112
1 2 F(x, 0.000186 5.29 3 10213 0 0
m, s 2)
Name /8042/03 04/21/99 09:19AM Plate # 0 NAAJ (SOA) pg 40 # 11
Table 3 Figure 16
Quantile Estimates for the ARCH Example Time Series and Mean-Excess Plots
of BMW Return Data
y 15 20 25 30
P(max{j i , . . . , j n} # y) 0.28229 0.65862 0.83794 0.91613
a New Hedging Tool for the Insurance Industry, Journal PUNTER, A. 1997. A Critical Evaluation of Progress to Date,
of Derivatives: 89104. in Practical Applications of Financial Market Tools to Cor-
CARDENAS, J., FRUCHARD, E., KOEHLER, E., MICHEL, C., AND THOM- porate Risk Management. International Risk Managment,
AZEAU, I. 1997. VaR: One Step Beyond, RISK 10(10):72 London: Emap Finance Ltd., pp. 712.
75. REISS, R.-D., AND THOMAS, M. 1997. Statistical Analysis of Extre-
DANIELSSON, J., HARTMANN, P., AND DE VRIES, C.G. 1998. The mal Values. Basel: Birkhauser.
Cost of Conservatism, RISK 11:1013. RESNICK, S.I. 1997. Discussion of the Danish Data on Large Fire
DOHERTY, N.A. 1997. Financial Innovation in the Management Insurance Losses, ASTIN Bulletin 27(2):13951.
of Catastrophe Risk, Joint Day Proceedings Volume of RESNICK, S.I., AND STARICA, C. 1996. Tail Index Estimation for
XXVIIIth International ASTIN Colloquium/7th International Dependent Data, available as TR1174.ps.Z at http://
AFIR Colloquium, pp. 126. www.orie.cornell.edu/trlist/trlist.html. To appear in Ann.
EMBRECHTS, P., KLUPPELBERG, C., AND MIKOSCH, T. 1997. Model- Applied Probability.
ling Extremal Events for Insurance and Finance. Berlin: RESNICK, S.I., AND STARICA, C. 1997. Smoothing the Hill Esti-
Springer-Verlag. mator, Advanced Applied Probability 29:27193.
EMBRECHTS, P., RESNICK, S.I., AND SAMORODNITSKY, G. 1998. ROOTZEN, H., AND TAJVIDI, N. 1997. Extreme Value Statistics
Living on the Edge, RISK 11(1):96100. and Windstorm Losses: A Case Study, Scandinavian Ac-
FREY, R., AND MCNEIL, A. 1998. Estimation of Tail-Related Risk tuarial Journal 1:7094.
Measures for Heteroscedastic Financial Time Series: an Ex- SCHMOCK, U. 1997. Estimating the Value of the Wincat Cou-
treme Value Approach. Zurich: ETHZ preprint. pons of the Winterthur Insurance Convertible Bond, Joint
GARMAN, M. 1997. Taking VaR to Pieces, RISK 10(10):7071. Day Proceedings Volume of XXVIIIth International ASTIN
GOLDIE, C.M. 1991. Implicit Renewal Theory and Tails of So- Colloquium/7th International AFIR Colloquium, pp. 231
lutions of Random Equations, Annals of Applied Proba- 59.
bility 1:12666. SIGMA. 1995. Natural Catastrophes and Major Losses in 1994:
HAAN, L. DE, RESNICK, S.I., ROOTZEN, H., AND DE VRIES, C.G. 1989. Third Highest Loss Burden in the History of Insurance,
Extremal Behaviour of Solutions to a Stochastic Differ- Swiss Reinsurance Company.
ence Equation with Applications to ARCH Processes, Sto- SIGMA. 1997. Natural Catastrophes and Major Losses in 1996:
chastic Processes and Their Applications 32:21324. High Losses from Man-Made Disasters, but Not Extremely
HOOGHIEMSTRA, G., AND MEESTER, L. 1995. Computing the Ex- Costly Losses from Natural Catastrophes, Swiss Reinsur-
tremal Index of Special Markov Chains and Queues, Sto- ance Company.
chastic Processes and Their Applications 65:17185. SMITH, R.L. 1990. Extreme Value Theory, in Handbook of Ap-
JAFFE, D.M., AND RUSSELL, T. 1996. Catastrophe Insurance, Cap- plicable Mathematics, Supplement, edited by W. Leder-
ital Markets and Uninsurable Risk. Philadelphia: Financial mann. Chichester: Wiley, pp. 43772.
Institutions Center, The Wharton School, pp. 96112. TAYLOR, S.J. 1986. Modelling Financial Time Series. Chichester:
KESTEN, H. 1973. Random Difference Equations and Renewal Wiley.
Theory for Products of Random Matrices, Acta Mathe- TILLEY, J.A. 1997. The Securitization of Catastrophic Property
matica 131:20748. Risks, Joint Day Proceedings Volume of XXVIIIth Inter-
LONGIN, F.M. 1997a. From Value at Risk to Stress Testing: The national ASTIN Colloquium/7th International AFIR Collo-
Extreme Value Approach. Paris: ESSEC and CEPR, quium, pp. 2753.
preprint. Value at Risk for End-Users. 1997. Risk Magazine, Supple-
LONGIN, F.M. 1997b. Beyond the VaR. Paris: ESSEC and CEPR, ment, March 1997.
preprint.
MCNEIL, A.J. 1997. Estimating the Tails of Loss Severity Dis- Discussions on this paper can be submitted until Oc-
tributions Using Extreme Value Theory, Astin Bulletin tober 1, 1999. The authors reserve the right to reply to
27(2):11737. any discussion. See the Submission Guidelines for Au-
MCNEIL, A.J. 1998. On Extremes and Crashes, RISK 11:99.
thors for detailed instructions on the submission of
MCNEIL, A.J., AND SALADIN, T. 1997. The Peaks over Threshold
discussions.
Method for Estimating High Quantiles of Loss Distribu-
tions, Proceedings of the XXVIIIth International ASTIN
Colloquium, pp. 2343.