Documente Academic
Documente Profesional
Documente Cultură
AUGUST 1, 2000
PHD THESIS
SUBMITTED TO THE FACULTY OF SCIENCE, UNIVERSITY OF
AARHUS
Prefa
e
This thesis is the out
ome of my PhD study and develops around
se
uritization of insuran
e risks. In parti
ular it
onsiders produ
ts
related to
atastrophe insuran
e. The thesis is split into three major
se
tions. The rst se
tion gives a short summary of ea
h of the ve
manus
ripts in
luded in the thesis, there is one in English and one
in Danish. The se
ond se
tion gives a longer review of the new main
results
ontained in the thesis and dis
usses how these results relates to
existing theories in the literature. The remainder of the thesis
onsists
of the ve manus
ripts. Cyni
s would say that the se
ond part exists
so that you will not have to read the manus
ripts.
Some of the work in the thesis was developed while I was visiting the
nan
ial and insuran
e mathemati
s group at the ETH, Zuri
h 1998-
1999 and Laboratory of A
tuarial Mathemati
s, University of Copen-
hagen 2000. I am grateful to Paul Embre
hts and Ragnar Norberg
respe
tively for making these stays possible.
The work in this thesis has beneted from stimulating dis
ussions,
suggestions and
omments from a number of persons. Credits are due
to Susan Bla
k (PCS), Amy Casey (CBoT), Bent Jesper Christensen,
Sam Cox, Freddy Delbaen, Paul Embre
hts, Damir Filipovi
, Asbjrn
Trolle Hansen, Jrgen Homann-Jrgensen, Dena Karras (CBoT), Ran-
di Mosegaard, Thomas Mller, Jrgen Aase Nielsen, Ragnar Norberg,
Jesper Lund Pedersen, Rolf Poulsen, Tina H. Rydberg, Uwe S
hmo
k,
Mogens Steensen, Mi
hael K. Srensen, Jim Welsh (PCS), and of
ourse my supervisor Hanspeter S
hmidli.
Contents
Summaries of Manus
ripts in the Thesis (English) i
Summaries of Manus
ripts in the Thesis (Danish) iv
1. Introdu
tion 1
2. The dieren
e between nan
ial and a
tuarial pri
ing 1
2.1. Insuran
e pri
ing 1
2.2. Pri
ing in nan
e 3
2.3. The interse
tion between insuran
e and nan
e 4
3. Se
uritization, the CAT-future 7
3.1. Des
ription of the CAT-future 8
3.2. The CAT-future pri
ing problem 9
3.3. Pri
ing based on a
tually reported
laims 10
4. Se
uritization, the PCS-option 13
4.1. Des
ription of the PCS-option 14
4.2. Why the PCS-option was an improvement 14
4.3. New pri
ing models for the PCS-option 16
5. Implied loss distributions 21
5.1. The models 22
5.2. The obje
tive fun
tion 26
5.3. The parameter estimation 28
5.4. Other relevant referen
es 31
6. The future of global reinsuran
e 32
6.1. The model 33
6.2. How to handle unknown risk in a
omplete market 34
6.3. The restri
ted premium
ase 35
6.4. The in
omplete market
ase 36
7. Con
lusion 37
Referen
es 40
Manus
ripts 44
Manus
ripts
Paper III: A new model for pri
ing
atastrophe insuran
e derivatives,
14 pages.
Working Papers
\The PCS-option, an improvement of the CAT-future", (Manus
ript,
University of Aarhus).
In 1992, Chi
ago Board of Trade (CBoT) introdu
ed the CAT-future
as an alternative to
atastrophe reinsuran
e. But the produ
t never
be
ame very popular, so in 1995 it was repla
ed by a new produ
t the
PCS-option. In relation to my PhD proje
t, I therefore started
ol-
le
ting information about the PCS-option. The basi
information was
obtained by reading [14℄ and [55℄. But some was also re
eived by mail-
ing with people from the PCS and the CBOT. After having re
eived
this information it seemed natural to gather it in a paper. This is done
here, together with an explanation of why the CAT-future was repla
ed
by the PCS-option, and why the PCS-option is an improvement. The
paper also explains how to hedge
atastrophe risk with PCS-options
and it
ompares the PCS-option with traditional reinsuran
e.
\A new model for pri
ing
atastrophe insuran
e derivatives", (Work-
ing Paper Series No. 28, Centre for Analyti
al Finan
e.)
Sin
e the introdu
tion of the insuran
e derivatives in 1992, it has been
a problem how to pri
e these produ
ts. The two main problems have
been the following. First if we
hoose a realisti
model for the under-
lying loss pro
ess the market will be in
omplete and there will exist
many equivalent martingale measures. Hen
e there will exist several
arbitrage free pri
es of the produ
t. Se
ond we want a Pareto like tail
for the underlying loss index, but heavy tails often give
omputational
problems. It is therefore natural to look for a model that solves both
these problems. In this paper we present a model whi
h in some sense
takes
are of both problems. The model is inspired by results from
Gerber and Shiu [43℄. In [43℄ it is shown that the Ess
her transform
is a unique and transparent te
hnique for valuing derivative se
urities
if the logarithms of the underlying pro
ess are governed by a
ertain
sto
hasti
pro
ess with stationary and independent in
rements (a Levy
pro
ess). In this paper we propose su
h a model, and by way of exam-
ple we
al
ulate pri
es for the PCS-option, but the approa
h
an also
be used for pri
ing other se
urities relying on a
atastrophe loss index.
\Implied loss distributions for
atastrophe insuran
e derivatives",
(Manus
ript, University of Aarhus).
In this paper we also pri
e
atastrophe insuran
e derivatives, but here
we lead our analysis in another dire
tion than in the previous papers.
We follow a pro
edure familiar to the
onventional option market whi
h
also is suggested by Lane and Mov
han in [46℄, namely rather than
Summaries iii
A
epterede artikler
Christensen, Claus Vorm and Hanspeter S
hmidli (1998), "Prisfast-
sttelse af katastrofeforsikrings produkter baseret p
a aktuelle rapporterede
skader", (udkommer i Insuran
e, Mathemati
s and E
onomi
s).
Resume: Denne artikel omhandler problemer ved prisfaststtelse
af et nansielt aktiv, som beror p a et index af rapporterede skader
fra katastrofeforsikringer. Problemet ved at prisfaststte s
adanne pro-
dukt er, at man p a et givet tidspunkt i handelsperioden ikke kender
det totale skadesbelb fra allerede indtrufne katastrofer. Man skal der-
for prisfaststte produktet udelukkende ud fra kendskabet til belbet
af de rapporterede skader til et givet tidspunkt. Denne artikel anviser
en m ade hvorpa man kan lse dette problem. Hovedideen i artiklen
er at modellere de samlede skader fra en katastrofe som en sammen-
sat (blandet) Poisson model. Vi opn ar derved muligheden for at se-
parere de individuelle skader og modellere rapporteringstidspunkterne
for skaderne. Denne nye model og en illustration af hvordan man kan
beregne priser i modellen er hovedbudskabet i artiklen.
Summaries v
Working Papers
\PCS-optionen, en forbedring af CAT-futuren", (Manuskript, Aarhus
Universitet).
I 1992 indfrte CBoT CAT-futuren som et alternativ til katastrofe gen-
forsikring. Men produktet blev aldrig rigtigt populrt, sa i 1995 blev
det erstattet af PCS-optionen. I relation til mit PhD studie, start-
ede jeg derfor med at indsamle information om PCS-optionen. Den
mest basale information stammer fra artiklerne [14℄ og [55℄. Derudover
blev en del information indhentet via personlig kontakt ved ansatte ved
PCS og CBoT. Efter at have indhentet denne information fandt jeg det
naturligt, at samle den i en artikel. Det er gjort i denne artikel, som
ogsa forklarer hvorfor CAT-futuren blev erstattet af PCS-optionen og
hvorfor PCS-optionen er en forbedring. Artiklen forklarer ogs a hvor-
dan man afdkker katastroferisiko og sammenligner PCS-optionen med
almindelig genforsikring.
\En ny model til prisfaststtelse af katastrofeforsikrings derivater",
(Working Paper Serie No. 28, Center for Analytisk Finansiering).
Siden nansielle katastrofe forsikringsprodukter blev indfrt i 1992, har
det vret et problem hvordan disse produkter skulle prisfaststtes.
Der har vret flgende to hovedproblemer. For det frste, hvis vi
vlger en realistisk model for den underliggende tabspro
es, s a bliver
markedet ufuldstndigt og som flge heraf eksisterer der mange kvi-
valente martingalm al. Dette betyder at der eksisterer mange arbitrage
frie priser p
a produktet. Det andet problem er at vi gerne vil have
at fordelingen for det underliggende tabs index har en tung hale, men
tunge haler giver ofte beregningsmssige problemer. Det har derfor
vret naturligt at lede efter en model der lser begge disse problemer.
I denne artikel prsenterer vi en model, der i en vis henseende tager
hjde for begge disse probelmer. Modellen er inspireret af resultater
fra Gerber og Shiu [43℄. I [43℄ er det vist at Ess
her transformen er en
entydig og umiddelbar teknik for vrdifaststtelse af aktiver, hvor log-
aritmen af den underliggende pro
es er styret af en bestemt stokastisk
pro
es med stationre og uafhngige tilvkster (en Levy pro
es). I
denne artikel foresl
ar vi en s
adan model, og som eksempel beregner vi
prisen for PCS-optionen. Metoden kan ogs a benyttes til prisfastst-
telse af andre aktiver, der beror pa et katastrofetabs index.
\Markeds udledte skadesfordelinger for katastrofeforsikrings derivater",
(Manuskript, Aarhus Universitet).
I denne artikel forsger vi ogs
a at prisfaststte katastrofe forsikrings
derivater, men vi drejer nu vores analyse i en anden retning end i
vi Summaries
1. Introdu
tion
During the nineties a highly dis
ussed theme among a
ademi
s has
been the interplay between insuran
e and nan
e. Some of the gen-
eral issues have been: the in
reasing
ollaboration between insuran
e
ompanies and banks; the dis
ussions about risk management method-
ologies for nan
ial institutions, and the emergen
e of nan
e related
insuran
e produ
ts, e.g.
atastrophe futures and options, PCS options,
index-linked poli
ies,........
This thesis develops around the interplay between insuran
e and -
nan
e and espe
ially around the pri
ing of nan
e related insuran
e
produ
ts (insuran
e derivatives). The insuran
e derivative was devel-
oped as a nan
ial produ
t whi
h should work as an alternative or
repla
ement of reinsuran
e. This meant that
ompanies that would
normally redu
e their risk by reinsuran
e,
ould now
onsider these
new nan
ial produ
ts as alternatives. One of the main dieren
es be-
tween the traditional
on
ept of reinsuran
e and these new produ
ts,
is the way they are pri
ed. Reinsuran
e
ontra
ts are pri
ed using
traditional a
tuarial methods, whereas derivatives should be pri
ed by
nan
ial methods of no arbitrage. To give an impression of the dif-
feren
es between these two methods of pri
ing, we start this se
ond
major se
tion with a des
ription of the two methods of pri
ing and
their intera
tion.
The rest of this se
ond se
tion of the thesis, gives a
hronologi
al
des
ription of how the market for
atastrophe insuran
e produ
ts has
developed, with a spe
ial fo
us on the pri
ing approa
h. Here rele-
vant literature is des
ribed and it is explained how my work relates to,
ontributes to, and extends various elds.
2.1. Insuran
e pri
ing. Let the annual premium for a
ertain risk
be , denote by Xi the losses in year i and assume that the Xi 's are
independent and identi
ally distributed. The
ompany has a
ertain
2 Pri
ing
From table 3.1 it is seen that all 10 events have happened during
the se
ond half of the period. One therefore gets the impression that
the frequen
y and severity of large losses has in
reased, whi
h is also
onrmed in [58℄. This in
rease is due to higher population densities,
more insured values in endangered areas and higher
on
entration of
values in industrialized
ountries.
The insuran
e industry already got the impression of this in
rease
in the early nineties after hurri
ane Andrew and the Northridge earth-
quake. And they soon realized that the reinsuran
e industry might la
k
the
apital to
over the huge
atastrophes of the future. To solve this
problem se
uritization of
atastrophe risk was invented. The idea of
se
uritization was to transfer some of the risk from the insuran
e mar-
ket to the nan
ial market, where the risk-bearing
apa
ity is mu
h
larger. This transfer should then be done by use of nan
ial instru-
ments su
h as options and futures on indi
es of
atastrophe losses or
atastrophe bonds. Catastrophe bonds are bonds where the payment
of the
oupon and/or the return of the prin
ipal of the bond is linked
to the non-o
urren
e of a spe
ied
atastrophi
event. In the rest
of the thesis we will denote these nan
ial instruments as
atastrophe
insuran
e derivatives.
We now give a short des
ription of one of the rst produ
ts of this
kind, namely the CAT-future introdu
ed by CBoT in 1992.
8 Se
uritization, the CAT-future
where IT is the ISO-index at the end of the reporting period, i.e. the
2
ratio between the losses in
urred during the event quarter and reported
up till three months later and the premium volume for the
ompanies
parti
ipating in the pool.
Example 3.1. The June
ontra
t
overs losses from events o
urring
in April, May and June and are reported to the parti
ipating
ompanies
by the end of September. The June
ontra
t expires on January 5th
the following year. The
ontra
t is illustrated by Figure 3.1.
Se
uritization, the CAT-future 9
Apr May June July Aug Sep Oct Nov Dec Jan
REPORTING PERIOD
3.3.2. The pri
ing of the CAT-future. Denoting by F~L (; t) the distri-
bution fun
tion of LT Lt under Q
onditioned on Ft . We
an then
2
express the pri
e at time t of the CAT-future as
EQ [LT ^ 2jFt ℄
2
=
(Lt + EQ [(LT 2
Lt ) ((LT 2
Lt ) (2 Lt ))+ j Ft ℄)
Z 1
(3.2) =
Lt + EQ[(LT 2
Lt ) j Ft ℄ (1 F~L (x; t)) dx :
2 Lt
The problem with the above expression, however, is that we have to
nd the n-fold
onvolutions of FD [T2 ~℄ (~ is uniformly distributed on
(t; T1 ), in order to
al
ulate the last term. To nd an expli
it expression
seems to be hard.
Histori
al data show that so far, the
ap 2 in the denition of the
CAT-future has not been rea
hed. The largest loss ratio was hurri-
ane Andrew with L1 = 1:79. Under the measure P we have that
fLT > 2g is a rare event. Sin
e we are dealing with
atastrophe
2
insuran
e, the market risk aversion
oeÆ
ient
annot be large. Oth-
erwise
atastrophe insuran
e would not be possible. We therefore as-
sume that fLT > 2g is also a rare event with respe
t to the mea-
2
sure Q, see also [32℄. The light tail approximation to our model then
assures that the tail of F~L (; t) is exponentially de
reasing. That is
R 1 ~
2 Lt (1 FL (x; t)) dx will be small as long as Q(LT > 2) is small.
2
The latter of
ourse depending on the risk aversion
oeÆ
ient , be-
ing small in order to be able to negle
t the last term. As in [32℄ we
therefore propose the approximation
(Lt + EQ [(LT Lt ) j Ft ℄) to the
2
pri
e of the CAT-future. For the exa
t
al
ulation of this expression,
see [17℄. The nal result is also stated here in Theorem 3.2.
Se
uritization, the PCS-option 13
In the extended model we assume that the i 's are sto
hasti
. This
an be seen as a measure of the severity of the
atastrophe. For sim-
pli
ity of the model, we assume that i
an be observed via reported
laims only. Of
ourse, in reality other information as TV-pi
tures or
reports from the ae
ted area will be available. Then for
laims o
ur-
ring before t we have some information on the intensity parameter i .
We therefore have to work with the posterior distribution of i given
Ft. It would be desirable if the prior and the posterior distribution
would belong to the same
lass, see the dis
ussion in [21, Ch.10℄. We
therefore
hoose i to be distributed. Let i (
; ). Under these
assumptions we again
al
ulate an approximation for the pri
e of the
CAT-future, see [17℄ for an exa
t
al
ulation. The nal result is stated
here in Theorem 3.6.
The above results are both approximations, so it is relevant to ask
how good these approximations are. This question is also
onsidered
in [17℄. From equation 3.2 we know that the approximation error is
given by the following expression:
Z 1
(3.3)
(1 F~L (x; t)) dx
2 Lt
where F~L (; t) denotes the distribution fun
tion of LT Lt under Q
onditioned on Ft . The reason for omitting this term was that it is hard
2
was
ertainly more diÆ
ult for other insurers. This
reated an informa-
tion asymmetry whi
h was a potential fa
tor preventing people from
entering the market of CAT-futures. This problem was solved for the
PCS-option, be
ause PCS reports the PCS loss indi
es on ea
h CBoT
trading day (the index is only
hanged if there are new
atastrophes
or if the index is adjusted) and neither Ameri
an Insuran
e Servi
es
Group nor any person employed by Ameri
an Insuran
e Servi
es Group
will dis
lose any estimate of total insured losses following a
atastro-
phe to any person prior to its oÆ
ial publi
ation. This means that all
investors re
eive the same information at the same time. Thereby the
problem of asymmetri
information is eliminated.
Another problem was the Moral Hazard problem. A
ompany from
the pool
ould manipulate data by delaying the report of a big loss so
that it rst would be in
luded in the next reporting period and thereby
never ae
t the index. The
ompany's intension for doing so,
ould be
that the
ompany had agreed to a short position of a future
ontra
t1 .
That this possibility existed,
ould also have prevented people from
entering the market of CAT-futures. This problem was also solved by
PCS
ondu
ting surveys of the market, when they estimate the loss
indi
es. These surveys are
ondential and they are not used dire
tly
in the estimation of the indi
es. So it is extremely diÆ
ult for insuran
e
ompanies to ae
t the indi
es, and thereby the Moral Hazard Problem
was eliminated.
A more serious problem
ould o
ur due to the reporting period being
too short. If a late quarter
atastrophe o
urs and
laims are slow in
developing, then the nal
laims ratio for the purpose of de
iding the
future payo
ould be low relative to the a
tual nal
laims ratio. This
problem o
urred in the Mar
h 1994
ontra
t period, the period of the
Northridge earthquake. The settlement ratio was low and the
ontra
t
pay-o did not truly re
e
t the a
tual
laim loss. The
onstru
tion
of the PCS-option also solves this problem. The PCS index does not
dire
tly depend on a number of reported
laims and the time from the
end of the event period to the time the index is settled is also longer
for the PCS-option than it was for the CAT-future.
These problems being solved, was probably the main reason for the
PCS-options higher trading a
tivity
ompared to the CAT-future. But
the fa
t that the new produ
t was more logi
ally
onstru
ted than
the old one,
ould also have had an ee
t. Hereby we mean that a
To present su
h a model was the aim of the manus
ript \A new model
for pri
ing
atastrophe insuran
e derivatives" Christensen [18℄ and is
to our knowledge the rst paper with this agenda. We now give a short
presentation of this manus
ript.
4.3.1. The model. The model we present below is inspired by Gerber
and Shiu [43℄. In [43℄ they show how one
an obtain a risk neutral
Ess
her measure in a unique and transparent way if the logarithms of
the value of the underlying se
urity is a Levy pro
ess. The idea is now
to
hoose su
h a model.
Let Lt be the underlying loss index for a
atastrophe insuran
e deriv-
ative, [0; T1 ℄ be the loss period and [T1 ; T2 ℄ be the development period.
We then assume that Lt for all t 2 [0; T2 ℄ is des
ribed by
Lt = L0 exp(Xt )
where Xt is a Levy pro
ess, and L0 2 IR+ . We model Xt dierently
in the loss period and the development period, for a similar model see
[55℄. The question is then how to model Xt for the loss period and for
the development period.
For t 2 [0; T1 ℄ we will model Xt by a
ompound Poisson pro
ess
Nt
X
Xt = Yi 8t 2 [0; T1℄
i=1
where Nt is a Poisson pro
ess with a xed parameter 1 , and Yi is
exponentially distributed with parameter . We hereby obtain one of
the desired properties namely as mentioned above the heavy-tail for
Lt , e.g. when Xt Exp() then Lt L0 is Pa(; L0 ) distributed. But
as mentioned above this model is not
hosen be
ause it is the most
obvious one but be
ause it has a heavy-tail, allows for
u
tuation and
gives a the possibility to express the pri
e in
losed-form. Therefore,
the model has some disadvantages
ompared to a more natural model,
rstly the fa
t that late
atastrophes be
ome more severe than earlier
ones and se
ondly L0 = L0 > 0. For this reason this model should only
be used as a rst \
rude" approximation to the real world. We have
tried to work out these problems, but this seems to be hard.
We have to
hoose a model for Xt for the development period. We
know that the adjustments are done both upwards and downwards we
will therefore again des
ribe Xt as a
ompound Poisson pro
ess for
t 2 [T1 ; T2 ℄
N~t T1
X
Xt = XT + 1
Y~i
i=1
18 Se
uritization, the PCS-option
where N~t is a Poisson pro
ess with a xed parameter 2 and Y~i is
normally distributed (N(; )), where the most natural
hoi
e of
is = 0 (unbiased previous estimates). In order to use the results
from Gerber and Shiu [43℄ we need to assume that the pro
ess Xt for
t 2 [0; T1 ℄ is independent of the pro
ess Xt XT for t 2 [T1 ; T2 ℄. In
1
the real world one will expe
t some dependen
e but the assumption is
invariable in order to use the results form [43℄.
The value of the option at time t is then
C (t; Lt ) = exp( r(T2 t))E [C (T2 ; LT )jFt ℄
2
where r is the risk free interest rate and E is the mean value a
ording
to a risk neutral measure. Before we
an pro
eed further in the
al
u-
lation of the option pri
e we will have to
hoose a risk neutral measure.
This is done in the next se
tions.
4.3.2. The pri
ing of the PCS-option. This se
tion des
ribes how to
ompute a risk neutral measure using the Ess
her Transform. The
theory was introdu
ed by Gerber and Shiu [43℄. However we need to
make some adjustments in order to use their results in our
ontext.
Let Lt be the value of the PCS index at time t
(4.1) Lt = L0 exp(Xt ); 8t 0;
where Xt is a Levy pro
ess. Let M (z; t) be the moment generating
fun
tion dened by:
Z 1
(4.2) M (z; t) := E [exp(zXt )℄ = exp(zx)F (dx; t)
1
provided the integral is nite, where F denotes the distribution fun
tion
for Xt . Be
ause of the independent stationary in
rements we then have,
(see [34℄, se
tion IX.5) , that
(4.3) M (z; t) = (M (z; 1))t
For any h 2 IR the Ess
her-Transformation F (dx; t; h) is dened as:
exp(hx)F (dx; t)
(4.4) F (dx; t; h) = :
M (h; t)
From this transformed density we dene the Ess
her-transformed mo-
ment generating fun
tion as:
Z 1
M (z + h; t)
(4.5) M (z; t; h) = exp(zx)F (dx; t; h) = :
1 M (h; t)
Then it follows from (4.3) and (4.5) that
(4.6) M (z; t; h) = (M (z; 1; h))t
Se
uritization, the PCS-option 19
By (4.6) it follows that the
ondition for hl in the loss period is:
(4.8) exp(r + 1 ) = Ml (1; 1; hl )
where Ml denotes that it is the moment generating fun
tion with re-
spe
t to the distribution in the loss period. For the development period
the situation is similar, the model for Xt is just dierent, see [18℄ for
further details. The
ondition for hd in the development period is:
(4.9) exp(r + 2 ) = Md (1; 1; hd )
where Md denotes that it is the moment generating fun
tion with re-
spe
t to the distribution in the development period.
The Radon-Nikodym derivative for the risk neutral Ess
her measure
on the -algebra Ft
an now be
hara
terized
8
dQ < ehl X t
Ml (hl ;t) t 2 [0; T1 ℄
jF =
dP t h X h (X X )
: e l T1 e d t T1
Ml (hl ;T1 ) Md (hd ;t T1 ) t 2 [T1 ; T2 ℄
where hl and hd are given by equation 4.8 and equation 4.9 respe
tively.
Although there is more than one equivalent measure, the risk neutral
Ess
her measure provides a unique and transparent answer. In [18℄ the
measure is justied by looking at a representative investor with a power
utility fun
tion, see [18℄ for further details.
The above results is then used to
al
ulate the
on
rete risk neu-
tral Ess
her measures for both the loss period and the development
period, see [18℄ for further details. Let us now
onsider the PCS
all option with exer
ise pri
e A,
ap K and expiring date T2 . Let
v1 (t) := ln(A=(L0 exp Xt )) and v2 (t) = ln(K=(L0 exp Xt )). The value
of the option at time t 2 [T1 ; T2 ℄
an then be expressed as:
C (t; Lt ) = e r(T2 t) Lt exp(2 (T2 t)) F (v2 (t); T2 t; hd + 1)
F (v1 (t); T2 t; hd + 1) + K (1 F (v2 (t); T2 t; hd ))
(4.10) A(1 F (v1 (t); T2 t; h ))
d
If t 2 [0; T1 ℄ then the value of the
all option at time t is given by
C (t; Lt ) = e r(T t) E [C (T2 ; LT )jFt ℄
2
2
transformed pro
ess in the loss period, see [18℄ for further details. Here
we
al
ulate the exa
t pri
e of the PCS-option.
5. Implied loss distributions
As we have seen in the previous se
tion, it is hard to nd realisti
models for the loss index, whi
h also allows one to express
onsistent
pri
es in a
losed form.
In Christensen [19℄ we therefore lead our analysis in another dire
-
tion. We follow a pro
edure familiar to the
onventional option market
whi
h is also suggested by Lane and Mov
han in [46℄. Rather than esti-
mating volatilities and
al
ulate
onsistent pri
es using, say the Bla
k
S
holes model, they take the traded pri
es and extra
t the volatilities
onsistent with those pri
es, i.e. nd the implied volatility. We
an-
not use the exa
t same pro
edure on the insuran
e derivative market,
sin
e we are not able to
hara
terize the pri
e by a single parameter.
However we
an do something similar. We
an
hoose a model for
the implied loss distribution and then estimate the implied parameters
from the observed pri
es.
This analysis
an be used to evaluate
heapness and dearness among
dierent pri
es and insuran
e derivative produ
ts. We simply
al
ulate
implied pri
es from the implied loss distributions and
ompare them
to the observed pri
es. There are two main problems related to this
analysis. First what kind of distribution should be used for the implied
loss distribution, and se
ond, how should the involved parameters be
estimated. These two questions are answered in the manus
ript [19℄,
where we base our analysis on data for the PCS-option. We now give
a summary of this manus
ript.
5.1. The models. In [19℄ we
onsider six dierent models for the im-
plied losses whi
h are all presented in this se
tion. Five of these models
are new. Before we present these models we rst give a general des
rip-
tion of the pri
e for a PCS
all-spread expressed by the implied loss
distribution.
Consider now a PCS
all-spread expiring at time T with upper and
lower strike Ku and Kl , respe
tively. Let FeLT and feLT be the implied
distribution fun
tion and the implied density fun
tion for the aggregate
PCS loss index (LT ) at time T . The value of the PCS
all-spread at
time 0 is then given by
PKu;Kl (L0 ; 0) = Ee [min(max(LT Kl ; 0); Ku Kl )℄
Z Ku
= (x Kl )feLT (x)dx + (Ku Kl )(1 FeLT (Ku ))
Kl
22 Implied loss distributions
The question is now, how is this implied loss distribution of the PCS
index related to the real statisti
al distribution, i.e. the distribution
under the P -measure? The implied distribution for the losses and
the real distribution for the losses will in general not be the same.
But based on the dis
ussion in [19℄, we only use models whi
h
ould
reasonably be used to des
ribe the real losses, when we model the
implied losses. We now present the six models for the implied losses.
Model 1. The rst model
onsidered is similar to the one suggested
by Lane and Mov
han [46℄, namely a
ompound Poisson model with
gamma distributed
laims, i.e.
NT
X
LeT = Yi
i=1
where LeT is the implied losses, NT Pois() and Yi (
; ). The
ni
e thing about this model is that we know the nth
onvolution of the
Y 's (Y1 + : : : + Yn (n
; )). This fa
t makes
omputations very
simple. A disadvantage of the model is that the
laims are light-tailed,
whereas data give eviden
e that the distribution of the aggregate
laims
is heavy-tailed. In this model we
an only approximate a heavy-tail by
hoosing low values of
and .
It is important to have this model in the analysis in order to see how
the results from this model dier from the following new and more
ompli
ated models.
Model 2. Looking at the listed
all-spreads we see, that the one
with the lowest strikes is the 40/60
all-spread (see Table 5.1 below).
The bid and ask for this
all-spread is 12 and 15 respe
tively, whi
h
are relatively large values for a produ
t that has a maximum pay-out
of 20. These fa
ts
ould therefore indi
ate that the market expe
ts the
loss index to be above a given threshold K0 for sure. If this is true and
K0 > 40, then there is no market for a 20/40
all-spread, be
ause the
market will expe
t the
all-spread to be worth 20 for sure. Based on
these indi
ations we now make an extension of model 1,
NT
X
LeT = K0 + Yi
i=1
where LeT is the implied losses, K0 is a
onstant indi
ating the threshold
the market expe
ts the losses to be above for sure, NT Pois() and
Yi (
; ). A
ording to the bid of the 40/60
all-spread (12), we
will not allow K0 to be above 52 (40+12). A possible interpretation of
this model is to think of K0 as the mean value of the \normal"
laims
and of the
ompound Poisson pro
ess as a model of the ex
esses.
Implied loss distributions 23
where f n (x) denotes the density for the nth
onvolution of the Pareto
distribution.
The rst 4
onvolutions are then found by the Rwell-known general
formula for R the Lebesgue
R
onvolution (f 2 (x) = 0x f (x y )f (y )dy ,
f 3 (x) = 0x f (x y )( 0y f (y z )f (z )dz )dy , : : : ). By only taking the
rst 4
onvolutions in the sum it should be possible for a
omputer to
al
ulate the expression. And if (the average numberPof
atastrophes)
is small, the approximation is good be
ause the term 1 n
5 e =n! will
be small.
Model 5 This model is inspired by the volatility surfa
e models
trying to explain the volatility smile, i.e. models where the implied
volatility depends on the strike of the option. We
onstru
t a similar
model where the most explanatory parameter in the implied loss dis-
tribution is dependent on the strike. We assume that the implied loss
index is Pareto distributed, i.e.
L~ T Pa(; )
being just a s
ale parameter. The most explanatory parameter in
this loss distribution is the parameter. We therefore
hoose to be
the strike dependent parameter, i.e. we assume that is a fun
tion of
the strike ((K ) = f (K )). The estimation of the parameters is done
in four steps, be
ause we have to
hoose the fun
tion f rst. The four
steps are as follows:
1. We assume that = 0 , i.e. independent of the strike. We then
let LeT P a(0 ; 0 ) and estimate the parameters 0 and 0 .
2. We now keep xed as 0 and then for ea
h PCS
all spread with
strikes Kli and Kui , we estimate an i from the traded pri
e or the
bid/ask spread dependent of what is available. These values are
then plotted. A possible pi
ture
ould be the one given by gure
5.1.
3. From this plot we
hoose a fun
tion to des
ribe our , i.e. if we
hoose a fun
tion f with three parameters a, b and
, a; b;
2 R.
We
an now des
ribe by (K ) = f (a; b;
; K ). The implied loss
distribution for a PCS
all option with strike K is therefore given
by
LeT Pa((K ); )
where (K ) = f (a; b;
; K ).
4. Cal
ulate the theoreti
al pri
es and estimate the parameters.
Model 6. The models 3, 4 and 5
ould also be extended by in
luding
a threshold as was done for model 1 in model 2. But we will desist from
doing this, as model 3 and 5 will be over parameterized and model 4
Implied loss distributions 25
-
20 40 60 80 100 120 Strike
term 4. In this term we value the information from the bid and the ask
equally, i.e. we prefer the theoreti
al pri
e to be in the middle of the
bid/ask spread. We
ap the single terms in the sum at 1/4, be
ause if
Pith = Pibid or Pith = Piask the single term in the sum is equal to 1/4,
and if Pith > Pibid or Pith < Piask then it is punished in term 1 or 2. How
mu
h this fourth term should be valued
ompared to term 1 and 2 is
adjusted by term 3. Term 3 is a
onstant Æ1 and a term denoting the
average length of the spread. In agreement with the
omments above,
we thereby obtain that, if the average length of the spreads is small,
we weight Pith being in the middle less than if the average length of the
spreads is large.
The terms 5 and 6 are in
luded in order to se
ure that the theoreti
al
pri
es do not get too far away from the single bids or asks. By too far
away we mean that a theoreti
al pri
e is punished if it is lower than
50% of a single ask or higher than 200% of a single bid. By the term
Æ2 we are able to adjust how mu
h the fth term should be valued
ompared to the other terms.
5.3. The parameter estimation. In this se
tion we estimate the
parameters and evaluate the six models des
ribed above. Before we
start to estimate we rst present the data.
The data material that we are going to use for this analysis are the
pri
es for the National PCS
all-spreads announ
ed by the CBoT on
January 7th 1999.
The National PCS
all-spreads announ
ed by the CBoT on January
7th 1999 is given by Table 5.1.
The rst
hange in the underlying PCS index was made January
19th, where the index in
reased from 0 to 7.6. We have
hosen the
28 Implied loss distributions
data from January 7th be
ause the last
hanges in the bids and asks
before January 19th were made here. If we take data from dates after
January 19th we have to take the value of the index into a
ount. If we
onsider data from a time point t where the PCS index is greater than
0, some adjustments have to be made. The implied losses at expiration
time T
an, at time t, be written as LeT = (LeT Lt ) + Lt , where Lt
is a
onstant and LeT Lt is the implied losses in the period from t
to T . LeT Lt
an then be des
ribed by the same models as we used
to des
ribe LeT , but the parameters will probably be
hanged. Even
though we are looking at a model where LeT is a stationary pro
ess, we
annot expe
t the same parameters sin
e the PCS index is in
uen
ed
by some large seasonal ee
ts.
The parameters are found by minimizing the obje
tive fun
tion, with
Æ1 = 0:001 and Æ2 = 0:1. We return to the dis
ussion of these parame-
ters later. The obje
tive fun
tion is a fun
tion depending on a higher
dimensional variable (the dimension is given by the number of param-
eters in the model). We therefore
hoose to minimize it by using a
modi
ation of the method of steepest des
ent des
ribed by Broyden
see [8℄ and [35℄.
The parameters found by minimizing the obje
tive fun
tion, the
or-
responding mean values and varian
es for the implied losses and the
theoreti
al pri
es are listed in table 5.2, table 5.3 and table 5.4, respe
-
tively.
M1 M2 M3 M4 M5 M6
Mean value 74 90 77 96 (73;91) 139
Varian
e 6096 8453 6178 11652 1 1
Table 5.3. Mean value and varian
e of implied losses.
Implied loss distributions 29
5.4. Other relevant referen
es. Let us end this se
tion with a short
des
ription of some other interesting papers in relation to insuran
e
derivatives. There is a huge amount of literature on the subje
t, a lot
of it being non-mathemati
al, e.g. [11℄, [36℄, [57℄ and [60℄. From the
more mathemati
al papers let me shortly des
ribe the following three.
The future of global reinsuran
e 31
Firstly, Bro
kett, Cox and Smith, [7℄ use a more a
tuarial pri
ing
approa
h. They assume that the traders do not have
omplete infor-
mation about the underlying loss pro
ess, but only information about
a range of values for the loss pro
ess, i.e.
1 E [LT ℄ 2 and 12 V ar[LT ℄ 22
Based on this information they are then able to derive a range of pri
es
for the insuran
e derivatives, see [7℄ for further details.
Se
ond, Rasmussen, [50℄, develops along the line of S
hweizer, [56℄,
and uses the minimal martingale measure to pri
e the PCS option. In
[50℄ it is shown that the equivalent minimal martingale measure exists
and it is shown how one
an nd the fair hedging pri
e of a PCS option
by
hoosing the equivalent minimal martingale measure as the pri
ing
measure.
Finally, we mention the paper by S
hmo
k, [54℄. This paper
onsid-
ers
atastrophe bonds issued by Winterthur. Several dierent models
are presented in order to evaluate the value of the
oupons, and it is
shown how substantial the model risk, inherent in pri
ing su
h nan
ial
produ
ts, is.
6. The future of global reinsuran
e
In this se
tion we will take a look into the future of global reinsur-
an
e. The results in this se
tion are based on Christensen [20℄.
Risk related to natural phenomena su
h as various
atastrophes has
traditionally been distributed through the insuran
e and reinsuran
e
system. Insuran
e
ompanies a
umulate the risk of individual entities
and redistribute the risk to the global reinsuran
e industry. But, as
dis
ussed earlier, it will be insuÆ
ient to manage this risk in su
h a
way in the future. A new way to managing su
h risk or unknown risk
in general is
alled for.
When we talk about unknown risk, we refer to risk whose frequen
y
we do not know, i.e. there is more than one estimate of the frequen
y
of the risk. Examples of unknown risk are environmental health risk of
new and little known epidemi
s, or risk indu
ed by s
ienti
un
ertainty
in predi
ting the frequen
y and severity of
atastrophi
events.
The problems related to unknown risk was rst mentioned by Chi-
hilnisky and Heal in [15℄ (a non mathemati
al paper), where they
argued that unknown risk should be managed by using traditional in-
suran
e pra
ti
e and by trading in the se
urity market simultaneously.
In the arti
le [20℄ we
ontinue and extend the ideas from [15℄. The
main purpose is to build a mathemati
al model that is able to handle
these problems.
32 The future of global reinsuran
e
In the following we will des
ribe the mathemati
al model from [20℄
and explain how we extend the ideas from [15℄ by
onsidering both
omplete and in
omplete markets and by
onsidering the
ase where
the premium
harged by the insuran
e
ompany is restri
ted.
In [20℄ we
onsider a general model for an insuran
e
ompany, where
the
ompany fa
es n states of the world. For ea
h of these states the
insuran
e
ompany is able to estimate the frequen
y of the risk, but
the risk related to the states is unknown. We show how the
ompany
should handle this unknown risk. This is done by using the statisti
al
approa
h to handle the known risk, i.e. the risk related to a given state,
and by using the e
onomi
approa
h to handle the risk related to the
dierent states.
6.1. The model. Let S denote the state of the world. We make the
following assumptions:
There are n states denoted by fs1; : : : ; sng; S 2 fs1; : : : ; sng.
The probabilities
orresponding to the n states are known
n
X
P (S = si ) = pi ; i = 1; : : : ; n; pi = 1
i=1
Fi is known for all i 2 f1; : : : ; ng, where Fi denotes the distribu-
tion of the loss (L) of the insuran
e
ompanies, given the state is
i (LjfS = si g Fi ). Let Li = LjfS = si g.
If the insuran
e
ompany knows the state S , then the statisti
al
approa
h by adding a safety loading would work, i.e. if the insur-
an
e
ompany knew that S = si , it would be reasonable to
harge
the premium Pi given by
Pi = E [Li ℄ + Æi
where Æi is a safety loading
al
ulated by a standard premium
al
ulation prin
iple.
There exist n \state se
urities" traded on the n states. Se
urity
number j pays the amount
ij if the state is i. Let
i be the ve
tor
i = (
i1 ; : : : ;
in) and let C be the matrix given by
2
1 3
C = 4 ... 5
n
Let further
~j be the j th
olumn in C .
The market is
omplete, i.e. the n
olumns in C are linearly inde-
pendent.
The future of global reinsuran
e 33
The market for these se
urities is arbitrage free and there exists a
unique risk neutral measure. We denote the risk neutral probabili-
ties by q1 ; : : : ; qn , and let q be the ve
tor given by q = (q1 ; : : : ; qn ).
From basi
nan
e
ourses it is known that these risk neutral prob-
abilities
an be used to pri
e the state se
urities, i.e. the pri
e of
state se
urity number i is given by the dis
ounted value of q
~j .
There exists a risk free se
urity and for simpli
ity we assume that
the risk free interest rate is zero. This is no loss of generality,
sin
e we
an dis
ount all se
urities.
We now have a model where the insuran
e
ompany exa
tly knows
how they should handle the insuran
e risk if the state of the world is
known. But be
ause of the un
ertainty about the state of the world,
the general risk for the insuran
e
ompany be
omes unknown. In the
next se
tion we will show how the insuran
e
ompany is able to handle
this unknown risk.
6.2. How to handle unknown risk in a
omplete market. The
expe
ted loss for the insuran
e
ompany is given by
E [L℄ = p1 E [L1 ℄ + + pn E [Ln ℄
To
over these losses, the insuran
e
ompany has to
harge a premium
P . But
harging a premium is not enough, sin
e we obtain a safety
loading in state i given by Æ~i = P E [Li ℄ if the insuran
e
ompany
harges a premium P . The problem
onne
ted with this, is that we do
not obtain the desired safety loading. For some i's we have that Æi < Æ~i
whi
h means that the insuran
e
ompany has been over
harging. And
for some i's we have that Æi > Æ~i , whi
h means that the insuran
e
ompany has been under
harging, whi
h
ould lead to a dangerous
position. Before we solve this problem, we make the two following
denitions.
Denition 6.1. A trading strategy for the insuran
e
ompany is
dened as a ve
tor m = (m1 ; : : : ; mn )T where mi denotes how many
se
urities i the insuran
e
ompany buys.
Denition 6.2. An optimal trading strategy for the insuran
e
ompany is a
ostless trading strategy su
h that
(6.1) P +
i m E [Li ℄ = Æi 8i = 1; ; n:
The questions are now whether it is possible to obtain this optimal
strategy and if so, what premium should be
harged in order to obtain
it? These questions are answered in the following theorem.
34 The future of global reinsuran
e
least square solution. We would then just have to repla
e the equation
P0 = P1 q1 + + Pn qn with an equation that makes sure that the
pri
e of the least square portfolio is equal to P0 . How this is done is
des
ribed in [20℄, see [20℄ for further details.
7. Con
lusion
What have we done in this thesis? Or perhaps more a
urately:
What are the
ontributions of the manus
ripts in
luded?
Christensen and S
hmidli [17℄ present a model for insuran
e fu-
ture pri
ing, whi
h only relies on the information available. The prod-
u
ts are pri
ed solely from observing the reporting stream. Contrary
to the existing literature we model the reporting times expli
itly. We
thereby obtain a more realisti
model.
The results of this arti
le rely on an approximation of the exa
t future
pri
e. One therefore has to be
areful applying the results derived,
be
ause the results will be ina
urate if the
ap-probability (P (LT > 2
2)) or the risk aversion
oeÆ
ient is \too large".
This paper suggests two ways of approximating the approximation
error, the gamma approximation and the Edgeworth approximation.
It is shown that they are both useful in the determination of the error
level even though the gamma approximation seems to be the best.
Christensen [16℄ is a gathering of information about the PCS-
option. It explains why the PCS-option repla
ed the CAT-future and
how the PCS-option is an improvement. The paper also explains how
to hedge
atastrophe risk with PCS-options and it
ompares the PCS-
option with traditional reinsuran
e.
Christensen [18℄ derives a new model for pri
ing insuran
e deriva-
tives whi
h allows for heavy-tails, and also provides a unique pri
ing
measure. The model is obtained by modeling the logarithms of the loss
pro
ess as a
ompound Poisson pro
ess with exponential distributed
marks in the loss period and with normal distributed marks in the de-
velopment period. The pri
e is found by evaluating the future pay-out
of the insuran
e derivative under the risk neutral measure derived by
the Ess
her approa
h. In the arti
le the exa
t pri
e in the
ase of the
PCS-option is
al
ulated.
Christensen [19℄ analyses pri
es for
atastrophe insuran
e deriva-
tives by looking at the \implied loss distributions" embedded in the
traded pri
es. And it gives answers to the two main problems in this
analysis. First, what kind of distribution should be
hosen for the
implied losses and se
ond, how should the involved parameters be es-
timated?
38 Con
lusion
Referen
es
[1℄ Aase, K.K. (1994): An equilibrium model of
atastrophe insuran
e futures
ontra
ts. Preprint.
[2℄ Albre
ht, P., A. Konig, and H.D. S
hradin (1994): Katastrophenversi-
herungsterminges
hafte: Grundlagen und Anwendungen im Risikomanage-
ment von Versi
herungsunternehmungen. Manuskript Nr. 2 Institut fur Ver-
si
herungswissens
haft, Universitat Mannheim, 1994.
[3℄ Beauregard, R.A. and Fraleigh, J.B. (1990), Linear Algebra, 2nd Edition,
Addison-Wesley Publishing Company.
[4℄ Barfod, A.M. and D. Lando (1996): On derivative
ontra
ts on
atastrophe
losses. Preprint, University of Copenhagen.
[5℄ Bladt, M. and T.H. Rydberg (1997): An A
tuarial Approa
h to Option Pri
-
ing under the Physi
al Measure and without Market Assumptions. Resear
h
Reports No. 388, Department of Theoreti
al Statisti
s, University of Aarhus.
[6℄ Bremaud, P. (1980): Point Pro
esses and Queues, Martingale Dynami
s.
Springer-Verlag, New York.
[7℄ Bro
kett, P.L., S.H. Cox and J. S
hmidt (1997): Bonds on the pri
e of
atas-
trophe insuran
e options on future
ontra
ts. Pro
eedings of the 1995 Bowles
Symposium on Se
uritization of Insuran
e Risk, Georgia State University, At-
lanta. SOA Monograph M-FI97-1, p. 1-7.
[8℄ Broyden, C.G. (1970), The
onvergen
e
lass of double-rank minimization al-
gorithms, J. Inst. Math. Appl.
[9℄ Buhlmann, H. (1980): An e
onomi
premium prin
iple. ASTIN Bulletin 11
(1), 52-60.
[10℄ Buhlmann, H. (1984): The general e
onomi
premium prin
iple. ASTIN Bul-
letin 14 (1), 13-21.
[11℄ Chanter, M.S., J.B. Cole and R.L. Sandor (1996): Insuran
e Derivatives: A
New Asset Class for the Capital Market and A New Hedging Tool for the
Insuran
e Industry. The Journal of Derivatives - Winter 1996.
[12℄ Chanter, M.S. and J.B. Cole (1997): The Foundation and Evolution of the
Catastrophe Bond Market: Catastrophe Bonds. Global Reinsuran
e - Septem-
ber 1997.
[13℄ Chang, C.W., J.S.K. Chang and M. Yu (1996): Pri
ing Catastrophe Futures
Call Spreads: A Randomized Operational Time Approa
h. The Journal of Risk
and Insuran
e, 1996, Vol. 63, No. 4. 599-617.
[14℄ The Chi
ago Board of Trade (1995): A User's Guide, PCS-options.
[15℄ Chi
hilnisky, G. and Heal, G. (1998), Managing Unknown Risks, The future of
Global Reinsuran
e, The Journal of Portfolio Management, pp 85-91, summer
1998.
[16℄ Christensen, C.V. (1997): The PCS Option: an improvement of the CAT-
future. Manus
ript, University of Aarhus.
[17℄ Christensen, C.V. and H. S
hmidli (1998): Pri
ing
atastrophe insuran
e
produ
ts based on a
tually reported
laims. to appear in Insuran
e: Mathe-
mati
s and E
onomi
s.
[18℄ Christensen, C.V. (1999): A new model for pri
ing
atastrophe insuran
e
derivatives. Working paper Series No. 28, Centre for Analyti
al Finan
e.
[19℄ Christensen, C.V. (2000): Implied loss distributions for
atastrophe insuran
e
derivatives Working paper Series No. , Centre for Analyti
al Finan
e.
40 Referen
es
[20℄ Christensen, C.V. (2000): How to hedge unknown risk. Working paper Series
No. , Centre for Analyti
al Finan
e.
[21℄ Cox, D.R. and D.V. Hinkley (1974): Theoreti
al statisti
s. Chapman and
Hall.
[22℄ Cox, S.H. and H. Pedersen (1997): Catastrophe Risk Bonds. Paper presented
at the 32nd A
tuarial Resear
h Conferen
e, August 6 - 8, University of Calgary,
Alberta, Canada.
[23℄ Cox, S.H. and R.G. S
hweba
h (1992): Insuran
e Futures and Hedging Insur-
an
e Pri
e Risk. The Journal of Risk and Insuran
e Vol. LIX No. 4. 628-644.
[24℄ Cummins, J.D. and H. Geman (1993): An Asian option approa
h to the
valuation of insuran
e futures
ontra
ts. Review Futures Markets 13, 517-557.
[25℄ Cummins, J.D. and H. Geman (1995): Pri
ing
atastrophe insuran
e futures
and
all spreads: an arbitrage approa
h. J. Fixed In
ome 4, 46-57.
[26℄ D'Ar
y, S.P, V.G. Fran
e and R.W.Gorvett (1999): Pri
ing Catastrophe Risk:
Could CAT Futures have
oped with Andrew? 1999 Casualty A
tuarial So
iety
\Se
uritization of Risk" Dis
ussion Paper Program.
[27℄ Delbaen, F. and J. Haezendon
k (1989): A martingale approa
h to premium
al
ulation prin
iples in an arbitrage free market. Insuran
e: Mathemati
s and
E
onomi
s 8, 269-277.
[28℄ DuÆe (1996): Dynami
Asset Pri
ing Theory, 2nd edition. Prin
eton Uiver-
sity Press.
[29℄ Eberlein, E. and J. Ja
od (1997): On the range of option pri
es. Finan
e and
Sto
hasti
s 1, 131-140.
[30℄ Embre
hts, P. (1996): A
tuarial versus nan
ial pri
ing of insuran
e.
Preprint, ETH Zuri
h.
[31℄ Embre
hts, P., C. Kluppelberg and T. Mikos
h (1996): Modelling Extremal
Events for Insuran
e and Finan
e. Appli
ations of Mathemati
s 33, Springer-
Verlag, Berlin.
[32℄ Embre
hts, P. and S. Meister (1997): Pri
ing insuran
e derivatives, the
ase of
CAT-futures. Pro
eedings of the 1995 Bowles Symposium on Se
uritization of
Insuran
e Risk, Georgia State University, Atlanta. SOA Monograph M-FI97-1,
p. 15-26.
[33℄ Embre
hts, P., S. Resni
k and G. Samorodnitsky (1997): Living on the Edge.
Preprint ETH Zuri
h.
[34℄ Feller, W. (1971): An introdu
tion to Probability Theory and its appli
ations.
Wiley, Vol. 2. 2nd ed. New York.
[35℄ Fielding, K. (1970), Fun
tion minimization and linear sear
h, Communi
ation
of the ACM, vol 13 8, 509-510.
[36℄ Frott, K.A. (1999): The Evolving Market for Catastrophi
Event Risk. Risk
Management and Insuran
e Review, 1999, Vol. 2, No. 3, 1-28.
[37℄ Folmer, H. and M. S
hweizer (1989): Hedging of
ontingent
laims under in-
omplete information. Applied Sto
hasti
Analysis, eds. M.H.A. Davis and R.J.
Elliott, Gordon and Brea
h, London.
[38℄ Folmer, H. and D. Sondermann (1996): Hedging of non-redundant
ontingent
laims. W. Hildebrand and A. Mas-Collel, (Eds.), Contribution to Mathemat-
i
al E
onomi
s in Honor of Gerard Debreu, Elsevier North-Holland, Amster-
dam, pp. 205-223.
[39℄ Geman, H. and M. Yor (1997): Sto
hasti
time
hange in
atastrophe option
pri
ing. Insuran
e: Mathemati
s and E
onomi
s 21, 185-193.
Referen
es 41
[40℄ Gerber, H.U. (1979): An introdu
tion to mathemati
al risk theory. Huebner
Foundation Monographs, Philadelphia.
[41℄ Gerber, H.U. and E.S.W. Shiu (1994): Option pri
ing by Ess
her transforms.
Trans. So
. A
tuar. 46, 99-191.
[42℄ Gerber, H.U. and E.S.W. Shiu (1995): A
tuarial approa
h to option pri
ing.
In Cox, S.H. (Ed.) Pro
eedings of the 1st Bowles Symposium GSU Atlanta.
So
iety of A
tuaries. To appear.
[43℄ Gerber, H.U. and E.S.W. Shiu (1996): A
tuarial bridges to dynami
hedging
and option pri
ing. Insuran
e: Mathemati
s and E
onomi
s 18, 183-218.
[44℄ Kluppelberg, C. and T. Mikos
h (1997): Large deviations of heavy-tailed ran-
dom sums with appli
ation in insuran
e and nan
e. Journal of Applied Prob-
ability 34. 293-308.
[45℄ Lane, M. and Finn, J. (1997): The perfume of the premium, Sedwi
k Lane
Finan
ial, Pro
eedings of the 1995 Bowles Symposium on Se
uritization of
Insuran
e Risk, Georgia State University, Atlanta. SOA Monograph M-FI97-
1, p. 27-35.
[46℄ Lane, M. and Mov
han, O. (1998), The perfume of the premium II, Sedwi
k
Lane Finan
ial, Trade Notes.
[47℄ Meister, S. (1995): Contribution to mathemati
s of
atastrophe insuran
e fu-
tures. Diplomarbeit, ETH-Zuri
h.
[48℄ Niehaus, G. and S.V. Mann (1992): The Trading of Underwriting Risk: An
Analysis of Insuran
e Futures Contra
ts and Reinsuran
e. The Journal of Risk
and Insuran
e 59 No. 4. 601-627.
[49℄ O'Brien, T. (1997): Hedging strategies using
atastrophe insuran
e options.
Insuran
e: Mathemati
s and E
onomi
s 21, 153-162.
[50℄ Rasmussen, A.L.B. (1998): Pri
ing PCS Options in a Finan
ial setup - an
In
omplete Market Model, master thesis, Department of Operation Resear
h,
University of Aarhus.
[51℄ Resni
k, S.I. (1992): Adventures in sto
hasti
pro
esses. Birkhauser.
[52℄ Rolski, T., H. S
hmidli, V. S
hmidt and J.L. Teugels (1999): Sto
hasti
pro
esses for insuran
e and nan
e. Wiley, Chi
hester.
[53℄ S
hmidli, H. : Le
ture notes in Risk Theory. Manus
ript, University of
Aarhus.
[54℄ S
hmo
k, U. (1999), Estimating the value of the WINCAT
oupons of the
Winterthur insuran
e
onvertible bond: A study of the model risk, ASTIN
Bulletin 29 (1), 101-164.
[55℄ S
hradin, H. R. und M. Timpel (1996): Einsatz von Optionen auf den
PCS-S
hadenindex in der Risikosteuerung von Versi
herungsunternehmen.
Mannheimer Manuskripte zu Versi
herungsbetriebslehre, Finanzmanagement
und Risikotheorie, Nr. 72.
[56℄ S
hweizer, M. (1993): Approximating random variables by sto
hasti
integrals
and appli
ations in nan
ial mathemati
s, ETH, Zuri
h.
[57℄ Sigma (1996): Insuran
e derivatives and se
uritization: New hedging perspe
-
tive for the US
atastrophe insuran
e market? Sigma publi
ation No. 5, Swiss
Re, Zuri
h.
[58℄ Sigma (2000): Natural
atastrophes and man-made disasters in 1999: Storms
and earthquakes lead to the se
ond-highest losses in insuran
e history. Sigma
publi
ation No. 2, Swiss Re, Zuri
h.
42 Referen
es
Manus
ripts
Paper I: Pri
ing
atastrophe insuran
e produ
ts based on a
tually
reported
laims, 17 pages.
Paper III: A new model for pri
ing
atastrophe insuran
e derivatives,
15 pages.
Abstra
t. This arti
le deals with the problem of pri
ing a nan-
ial produ
t relying on an index of reported
laims from
atastro-
phe insuran
e. The problem of pri
ing su
h produ
ts is that, at a
xed time in the trading period, the total
laim amount from the
atastrophes o
urred is not known. Therefore one has to pri
e
these produ
ts solely from knowing the aggregate amount of the
reported
laims at the xed time point. This arti
le will propose a
way to handle this problem, and will thereby extend the existing
pri
ing models for produ
ts of this kind.
1. Introdu
tion
Modelling
laims from a
atastrophe a
tuaries use heavy tailed dis-
tributions, su
h as the Pareto distribution. This means that the aggre-
gate
laim basi
ally is determined by the largest
laim, see [8℄ or [13℄.
This ee
t be
ame
learly visible in the early 90's, when the insuran
e
industry had to
over huge aggregate
laims in
urring from
atastro-
phes. Be
ause
ertain
atastrophi
events like earthquakes, hurri
anes
or
ooding are typi
al for some areas, a properly
al
ulated annual
premium would be nearly as high as the loss insured. From an a
tu-
arial point of view, su
h events are not insurable. But people living in
su
h areas need prote
tion. One possibility would be the government
(tax payer) to take over the risk, as it is the
ase for
ooding in the
Netherlands. Another possibility are futures or options based on a loss
index. Here the risk is transfered to private investors. A des
ription of
these produ
ts
an be found for example in [2℄ or [14℄.
In 1992 the Chi
ago Board of Trade (CBoT) introdu
ed the CAT-
futures. This future is based on the ISO-index, whi
h measures the
amount of
laims o
ured in a
ertain period and reported to a par-
ti
ipating insuran
e
ompany until a
ertain time. The produ
t never
be
ame popular among private investors. The reasons were that the
index only was announ
ed on
e before the settlement date, there was
information asymmetry between insurers and investors, and that there
1991 Mathemati
s Subje
t Classi
ation. 62P05.
Key words and phrases. Insuran
e futures; Derivatives; Claims-pro
ess; Catas-
trophe insuran
e; Mixed Poisson model; Change of measure; Expe
ted utility;
Approximations.
1
2 C. VORM CHRISTENSEN AND H. SCHMIDLI
where IT is the ISO-index at the end of the reporting period, i.e. the
2
ratio between the losses in
urred during the event quarter and reported
up till three months later and the premium volume for the
ompanies
parti
ipating in the pool.
Example 1.1. The June
ontra
t
overs losses from events o
urring
in April, May and June and are reported to the parti
ipating
ompanies
by the end of September. The June
ontra
t expires on January 5th,
the following year. The
ontra
t is illustrated by Figure 1.1.
1.2. The CAT-future pri
ing problem. Cummins and Geman [7℄
were the rst to pri
e the insuran
e futures. Their approa
h was quite
4 C. VORM CHRISTENSEN AND H. SCHMIDLI
Apr May June July Aug Sep Oct Nov Dec Jan
REPORTING PERIOD
dierent from the approa
h used in this paper. As model they used
integrated geometri
Brownian motion. This allowed them to apply
te
hniques arising from pri
ing Asian options. The model, however,
seems to be far from reality. At times where a
atastrophe o
urs or
shortly thereafter, one would expe
t a strong in
rease of the loss index.
It therefore is preferable to use a marked point pro
ess as it is popular
in a
tuarial mathemati
s.
The pri
e to pay for the more realisti
model is \non-uniqueness" of
the market, see [1℄ and [9℄ for further details. In fa
t, the index (It ) is
not a traded asset. Thus markets
annot be
omplete. Moreover, as it
is the
ase for term stru
ture models, any equivalent measure may be
used for no-arbitrage pri
ing. However, the preferen
es of the agents
in the market will determine whi
h martingale measure applies.
In this arti
le we follow the approa
h of Embre
hts and Meister [9℄.
There the general equilibrium approa
h is used, where all the agent's
utility fun
tions are of exponential type. More pre
isely, let Ft be the
pri
e of the future, Lt the value of the losses o
ured in the event quar-
ter and reported till time t, Ft the information at time t, premiums
earned and let
= 25 000=. Then the pri
e at time t, is (see [9, p.19℄)
E [exp(L1 ) (LT ^ 2) j Ft ℄
Ft =
P :
2
(1.2)
EP [exp(L1) j Ft ℄
In parti
ular, EP [exp(L1 )℄ has to exist. The market will determine
the risk aversion
oeÆ
ient .
The term exp(L1 )=EP [exp(L1)℄ is stri
tly positive and integrates
to one. Thus it is the Radon-Nikodym derivative dQ=dP of an equiv-
alent measure. In the spe
i
model we will
onsider, the pro
ess (Lt )
follows under to the measure Q the same model (but with dierent pa-
rameters) as under P . We will use this fa
t to
al
ulate the pri
e of the
CAT-future and the PCS-option. This
hange of measure is similar to
the Ess
her method des
ribed in [11℄ and [14℄. If we assume that pro-
portional reinsuran
e is possible, the premiums are fairly split between
insurer and reinsurer, and that the proportion held in the portfolio
an
be
hanged at any time, then the index (L1 (T ) (T )) would be
ome
a traded asset, where L1 (T ) are all the
laims o
ured till time T and
PRICING CATASTROPHE INSURANCE PRODUCTS 5
(T ) are the premiums earned to
over the
laims o
uring till time T .
In our model (T ) would be a linear fun
tion. This would imply that
the pro
ess (L1 (T ) (T ) : T 0) is a martingale under the pri
ing
measure. This
ondition will determine the risk aversion
oeÆ
ient ,
see for instan
e [15℄.
To pro
eed further in the
al
ulation of the future pri
e, one has to
hoose a model for (Lt ). [1℄ used a
ompound Poisson model. This
an be seen as
atastrophes o
urring at
ertain times and
laims are
reported immediately. In su
h a model there would not be a need
for the prolonged reporting period. In [9℄ a doubly sto
hasti
Poisson
model is introdu
ed. Here, a high intensity level will o
ur shortly after
a
atastrophe, where more
laims are expe
ted to be reported. In [12,
Example 5.3℄ the asymptoti
expe
ted value and asymptoti
varian
e
for a general
ompound pro
ess are obtained.
The aim of this paper is to model the
laims reported to the
om-
panies as individual
laims with a reporting lag. This is done by mod-
elling the aggregate
laim from a single
atastrophe as a
ompound
(mixed) Poisson model. We thereby obtain the possibility to separate
the individual
laims and to model the reporting times of the
laims.
In Se
tion 3.1 we
al
ulate the future pri
e using a
ompound Pois-
son model, whereas in Se
tion 3.2 the results are extended by using a
ompound negative binomial model, represented as a mixed
ompound
Poisson model. We thereby
an estimate the mixing parameter from
the reporting
ow.
For the rest of this se
tion we work with the measure P
onditioned
on Ft . Let
i (T2 )
MX
Si = Yi:j :
j =Mi (t)+1
Nt +1 ; : : P
1
: ; NT , is
ompound Poisson distributed with intensity pa-
1
distribution with moment generating fun
tion expf(mY (r) 1)g. This
yields
EP [expfL1 g℄ = expf(e(mY () 1) 1)g :
Let us
onsider now the pro
ess (Lt ) under the measure Q. For an
introdu
tion to
hange of measure methods we refer to [13℄. A simple
al
ulation yields that under Q the pro
ess (Lt ) is of the same type,
8 C. VORM CHRISTENSEN AND H. SCHMIDLI
only with dierent parameters. (Nt ) is a Poisson pro
ess with rate
h M1
nX oi
~ = EP exp Y 1j = expf(mY () 1)g :
j =1
The number of
laims of the i-th
atastrophe is Poisson distributed
with parameter ~ = (mY (R)) and the individual
laims have the dis-
tribution fun
tion F~Y (x) = 0x ey dFY (y )=mY (). The lags Dij have
the same distribution as under P .
The pri
e of a CAT-future is therefore
EQ [LT ^ 2℄. Denoting the
2
Then papprox
t for a given risk aversion
oeÆ
ient is given by
Nt
25 000 X
Lt + (FD (T2 i ) FD (t i ))
i=1
~ T1
+( t)EQ [FD (T2 ~ Q [Y ℄ :
~)℄ E
for t 2 [0; T1 ℄ and
NT1
25 000 X
~ Q [Y ℄
Lt + (FD (T2 i ) FD (t i ))E
i=1
for t 2 [T1 ; T2 ℄.
Proof: We only
onsider EQ [LT Lt 2 j Ft℄. From the
onsiderations
in Se
tion 2 we know that for t < T1
EQ [(LT Lt ) j Ft ℄
2
Nt
hX i
= EQ ~ (FD (T2
i ) FD (t i )) Nt ; 1 ; : : : ; Nt
i=1
h XNT1 i
+ EQ ~ D (T2
F ~i ) EQ [Yij ℄
i=Nt +1
Nt
X
= (FD (T2 i ) FD (t i ))
i=1
~ T1
+( t)EQ [FD (T2 ~ Q[Y ℄ :
~)℄ E (3.3)
Note that
Z T2 t
1
EQ [FD (T2 ~)℄ = EP [FD (T2 ~)℄ = FD (s) ds ; (3.4)
T1 t T2 T1
VarQ[LT Lt ℄ 2
6 24 72
The inverse of expfr2 =2g is easily found to be the normal distribution
fun
tion (x). For the other terms we derive
Z 1 Z 1
rner =2 = (erx )(n) 0 (x) dx =( 1)n erx (n+1) (x) dx :
2
1 1
Thus the inverse of rner =2 is ( 1)n times the n-th derivative of . The
2
approximation yields
P [LT
2
Lt x℄ = P [Z z ℄
a2
(z) a63 (3) (z) + 24
a4 (4)
(z ) + 3 (6) (z )
72
PRICING CATASTROPHE INSURANCE PRODUCTS 13
p
where z = (x E [LT Lt ℄)= Var[LT
2 2
Lt ℄. The approximation error
therefore is approximated by
p
AP(E) =
Var[LT Lt ℄
2
Z 1
a3 (3) a a2
(z ) (z ) + 4 (4) (z ) + 3 (6) (z ) dz :
z
0
6 24 72
p
where z0 = ((2 Lt ) E [LT Lt ℄)= Var[LT Lt ℄.
2 2
We now have
onstru
ted two ways of approximating the AE. The
question is then whether we obtain a better pri
e if we
orre
t the
un
apped future pri
e with these approximations, or we are better o
just using the un
apped future pri
e dire
tly? We will now look at an
example in order to answer this question.
4.2. Example. The
apped future pri
e (FtC ) is
al
ulated a
ording
to equation (1.2)
E [exp(L1 ) (LT ^ 2) j Ft ℄
FtC =
P 2
EP [exp(L1 ) j Ft ℄
where a reliable value of the expression is obtained by Monte-Carlo
simulations. In order to use MC we make the following assumptions:
The
laim sizes are exponentially distributed with parameter ,
The reporting lags are exponentially distributed with parameter
.
The un
apped future pri
e (FtU ) is
al
ulated a
ording to Theorem 3.2.
We will keep all the parameters xed in the example, ex
ept from the
premium and the risk aversion
oeÆ
ient in order to see how the
approximations depend on these two parameters. We use the following
parameters:
T1 = 1 T2 = 2 t = 0:5
=6 Nt = 3 = 0:0005
1 = 0:1 2 = 0:25 3 = 0:4
M1 (t) = 698 M2 (t) = 528 M3 (t) = 259
= (1 + )12 106 = 1000
Lt = E [Lt ℄ = 2:97 106 =3
is
al
ulated by the expe
ted value prin
iple with safety loading
under the assumption that all the
laims will be reported.
The parameters are
hosen su
h that P (LT > 2) is
onsistent with
2
the few data that we had. None out of the approximately 80 available
settlement values ex
eeded the level 2. For dates before 1992 the ISO
index had to be estimated from the nal aggregate loss value (L1 ).
The largest values of the ratio LT = there have been seen so far is
2
1.7893 (the Eastern Loss Ratio from Hurri
ane Andrew, Sept. 1992)
14 C. VORM CHRISTENSEN AND H. SCHMIDLI
and 1.0508 (the Western Loss Ratio from Northridge Earthquake, 2nd
Mar
h 1994). In our example with = 0:05 we have that P (LT = >
1:79) 0:01.
2
For dierent values of and , Table 4.1 shows the values of FtC ,
Ft , the approximation error AE = FtU FtC by using the un
apped
U
future pri
e, the approximation error AE(G) = (FtU AP(G)) FtC if
we
orre
t FtU by the gamma approximation to (4.1), and nally the
approximation error AE(E) = (FtU AP(E)) FtC if we
orre
t FtU by
the Edgeworth approximation (4.1).
From Table 4.1 we see, that for all the
hosen parameters the un-
apped future pri
e seems to approximate the
apped future pri
e fairly
well, and best when the risk aversion
oeÆ
ient is small or the safety
loading is large. But are the
hosen parameters reasonable?
Let us rst dis
uss the parameter. In insuran
e the safety loading
is always positive, and looking at real data the safety loading seems to
be \large" when we are
onsidering
atastrophe insuran
e. By \large"
we mean that the event fLT > 2g never has o
ured.
2
The parameter is the risk aversion
oeÆ
ient for the single in-
suran
e
ompany when pri
ing in a utility maximization framework
(see [9℄ for further details), or the markets risk aversion when pri
ing
in a general equilibrium model (see [9℄ for further details). The rst
thing to note on the parameter is that the parameter is pri
e de-
ned, i.e. it depends on the way we pri
e the losses. Here the values of
the losses are \large" and therefore the parameter be
omes \small".
The parameters are then
hosen in su
h a way that dierent pri
es are
represented, for = 0:15 and = 1 10 8 the
apped future pri
e is
PRICING CATASTROPHE INSURANCE PRODUCTS 15
5. Con
lusion
This paper develops a model for insuran
e future pri
ing, whi
h only
relies on the information available. The produ
ts are pri
ed solely from
observing the reporting stream. Contrary to the existing literature we
model the reporting times expli
itly. We thereby obtain a more realisti
model.
The results of this arti
le rely on an approximation to the exa
t
future pri
e. One therefore has to be
areful applying the results
derived, be
ause the results will be ina
urate if the
ap-probability
(P (LT > 2)) or the risk aversion
oeÆ
ient is \too large".
2
This paper suggest two ways to approximate the approximation er-
ror, the gamma approximation and the Edgeworth approximation. It
is shown that they both are useful in the determination of the error
level even though that the gamma approximation seems to be the best.
The results are derived spe
ially for the CAT-futures, even though
an improved nan
ial
atastrophe insuran
e produ
t, the PCS-option,
was introdu
ed in 1995. For a des
ription of the PCS-option and an
explanation of why the CAT-future was improved see [5℄. The results
16 C. VORM CHRISTENSEN AND H. SCHMIDLI
from this arti
le
annot dire
tly be used for pri
ing PCS-options be-
ause they have another stru
ture. But, be
ause of a strong
orrelation
between
laims reported and the PCS-index some of the ideas may be
used. This is a topi
for further resear
h.
A
knowledgement
The authors thank a referee for his
omments that lead to an im-
provement of the presentation.
Referen
es
1. Aase, K.K. (1994): \An equilibrium model of
atastrophe insuran
e futures
ontra
ts." Preprint.
2. Albre
ht, P., A. Konig, and H.D. S
hradin (1994): \Katastrophenversi-
herungsterminges
hafte: Grundlagen und Anwendungen im Risikomanage-
ment von Versi
herungsunternehmungen." Manuskript Nr. 2 Institut fur Ver-
si
herungswissens
haft, Universitat Mannheim, 1994.
3. Barfod, A.M. and D. Lando (1996): \On derivative
ontra
ts on
atastrophe
losses." Preprint University of Copenhagen.
4. Buhlmann, H. (1980): \An e
onomi
premium prin
iple." ASTIN Bulletin 11
(1), 52-60.
5. Christensen, C.V. (1997): \The PCS Option: an improvement of the CAT-
future." Manus
ript, University of Aarhus.
6. Cox, D.R. and D.V. Hinkley (1974): \Theoreti
al statisti
s." Chapman and
Hall.
7. Cummins, J.D. and H. Geman (1993): \An Asian option approa
h to the
valuation of insuran
e futures
ontra
ts." Review Futures Markets 13, 517-557.
8. Embre
hts, P., C. Kluppelberg and T. Mikos
h (1997): \Modelling extremal
events for insuran
e and nan
e." Appli
ations of Mathemati
s 33, Springer-
Verlag, Berlin.
9. Embre
hts, P. and S. Meister (1997): \Pri
ing insuran
e derivatives, the
ase of
CAT-futures." In: Se
uritization of Insuran
e Risk: 1995 Bowles Symposium.
SOA Monograph M-FI97-1, p. 15-26.
10. Gerber, H.U. (1979): \An introdu
tion to mathemati
al risk theory." Huebner
Foundation Monographs, Philadelphia.
11. Gerber, H.U. and E.S.W. Shiu (1994): \Option pri
ing by Ess
her transforms."
Transa
tions - So
iety of A
tuaries. 46, 99-191.
12. Kluppelberg, C. and T. Mikos
h (1997): \Large deviations of heavy-tailed ran-
dom sums with appli
ations in insuran
e and nan
e." Journal of Applied Prob-
ability 34, 293-308.
13. Rolski, T., H. S
hmidli, V. S
hmidt and J.L. Teugels (1999): \Sto
hasti
pro-
esses for insuran
e and nan
e." Wiley, Chi
hester.
14. S
hradin, H.R. and M. Timpel (1996): \Einsatz von Optionen auf den PCS-
S
hadenindex in der Risikosteuerung von Versi
herungsunternehmen." Mann-
heimer Manuskripte zu Versi
herungsbetriebslehre, Finanzmanagement und
Risikotheorie, Nr. 72.
15. Sondermann, D. (1988): \Reinsuran
e in arbitrage-free markets." Insuran
e:
Mathemati
s and E
onomi
s 10, 191-202.
PRICING CATASTROPHE INSURANCE PRODUCTS 17
1. Introdu
tion
The insuran
e industry has been hit very hardly in the 1990s by
their
atastrophe insuran
es. This has been
aused by a re
ord num-
ber of natural
atastrophe losses, of whi
h the insuran
e premiums only
overed a small part. At the same time many of the
atastrophe pre-
miums are very large (approximately the value of the maximal losses),
so there is hardly no room for in
reasing the
apa
ity in the insur-
an
e market. Examples of risks whi
h demand su
h large premiums
are
ooding in the Netherlands o
uring every spring and earthquakes
i L.A. also o
uring regularly. Thus the sear
h for new
apa
ity has
led to the prospe
t of trading insuran
e risk not only within the tradi-
tional insuran
e system but also transferring them to the more liquid
nan
ial markets.
On the De
ember 11, 1992 CBoT made the rst attempt to do so.
They laun
hed futures on
atastrophe loss indi
es and related options
(CAT-future and options). The CAT-option, also referred to as the
future option, has as underlying instrument one
atastrophe insuran
e
future
ontra
t. Be
ause of this relation the arti
le will only
onsider
the CAT-future in the se
tion where these old produ
t is under
on-
sideration). Following initial diÆ
ulties, whi
h will be explained later,
these standardized
ontra
ts have been improved, and on September
29, 1995 CBoT introdu
ed the PCS-options. The underlying assets
of the PCS-options are the PCS indi
es. These loss indi
es are pro-
vided daily to the CBoT by the Property Claim Servi
es (PCS), whi
h
is the re
ognized industry authority for
atastrophe property damage
estimates.
Date : April, 1998.
1
2 C. VORM CHRISTENSEN
This arti
le will rst give a des
ription of the PCS-option, then it will
des
ribe the ISO index whi
h was the underlying index of the CAT-
futures and also the main reason for the produ
t's problems. Then
the index for the PCS-options is des
ribed and it is explained how the
PCS-options improved the CAT-future. It will then be shown how to
hedge with PCS-options, and nally the arti
le gives a des
ription of
PCS-options versus reinsuran
e.
index value at the end of the
hosen development period will be used
for settlement purposes, even though PCS loss estimates may
ontinue
to
hange.
PCS-options settle in
ash on the last business day of the develop-
ment period. The settlement value (LT ) for ea
h index represents the
sum of then-
urrent PCS insured loss estimates provided and revised
over the loss and development periods. PCS-options are options of Eu-
ropean type, that means that they
an be exer
ised on the expiration
day at the end of the development period only.
The value of the PCS-
all option at expiration day T , exer
ise pri
e
X and
ap value K
an be expressed as
C (T; L(T )) = min(max(L(T ) X; 0); K X)
Due to diÆ
ulties in trading options in industry loss dollar amounts,
the CBoT has developed a pri
ing index to re
e
t dollar loss amounts
raging from $ 0 to $ 50 billion. Ea
h PCS loss index represents the
sum of then-
urrent PCS estimates for insured
atastrophi
losses in
the area and loss period
overed, divided by $ 100 million and rounded
to the nearest rst de
imal point. PCS-options pri
es or premiums,
are quoted in points and tenths of a point. Ea
h point equals $ 200;
ea
h tenth of a point equals $ 20. We will end this subse
tion with an
example.
Example 2.1. Let us
onsider a reinsurer who buys a June Eastern
small
ap
all PCS-option with strike value of 20 and a development
period of six-months. This
ontra
t tra
ks losses from
atastrophi
events o
uring in the Eastern region between April 1 and June 30
1997. The six-month development period runs from July 1 to De
ember
31. The option will thus settle on De
ember 31, 1997 a
ording to the
settlement value of the index.
So in this example the reinsurer re
eives $ 3140. If the loss index has
been estimated above $ 20 billion let us say $ 23 billion then the value
of the
all option would be
C (T; L(T )) = min(max(230 20; 0); 200 20) $200 = $36000
and the reinsurer would then (only) have re
eived $ 36000 be
ause it
was a small
ap option.
3. Loss estimation
As mentioned in the introdu
tion, the PCS-options is an improve-
ment of the CAT-futures. The main problems with CAT-futures were
aused by the underlying asset, the so
alled ISO index. The improve-
ments have therefore mainly been a
hieved by
hanging this underlying
asset. This subse
tion will highlight some of the problems of the CAT-
futures, and explain how the introdu
tion of the PCS-option solved
some of them. The information about the CAT-future is obtained
from [1℄ and [3℄. Let us rst
onsider the ISO index.
3.1. The ISO index. Ea
h quarter approximately 100 Ameri
an in-
suran
e
ompanies reported property loss data to the ISO (Insuran
e
Servi
e OÆ
e, a well known statisti
al agent). ISO then sele
ted a pool
of at least ten of these
ompanies on basis of size, diversity of business,
and quality of reported data. The ISO index was then
al
ulated as
the loss ratio of this pool.
The ISO index:
reported in
urred losses
ISO index = :
earned premiums
The list of
ompanies in
luded in the pool was announ
ed by the
CBoT prior to the beginning of the trading period for that
ontra
t.
The CBoT also announ
ed the premium volume of the
ompanies par-
ti
ipating in the pool prior to the start of the trading period. Thus
the premium in the pool was a known
onstant throughout the trad-
ing period, and pri
e
hanges were attributed solely to
hanges in the
markets expe
tation of loss liabilities.
CAT-futures were traded on a quarterly
y
le, with
ontra
t months
Mar
h, June, September, and De
ember. A
ontra
t for any given
alendar quarter (the event quarter) was based on losses o
uring in
the listed quarter, and beeing reported to the parti
ipating
ompanies
by the end of the following quarter. The six month period following the
start of the event quarter is known as the reporting period. The three
additional reporting months following the
lose of the event quarter are
to allow for loss settlement lags that are
ommon in insuran
e. The
ontra
ts expire on the fth day of the fourth month following the end
of the reporting period. The additional three months following the
reporting period is attributable to data pro
essing lags. Trading was
THE PCS-OPTION, AN IMPROVEMENT 5
ondu
ted from the date the
ontra
t was listed until the settlement
date.
Example 3.1. The June
ontra
t
overs losses from events o
uring
in April, May and June as reported to the parti
ipating
ompanies by
the end of September. The June
ontra
t expires on January 5th the
following year. The
ontra
t is illustrated by the gure below.
Apr May June July Aug Sep Oct Nov Dec Jan
REPORTING PERIOD
Finally the settlement value for the CAT-futures was given by:
FT = $25000 min(IT ; 2)
where IT is the ISO index at time T, i.e. the ratio between the losses
in
urred during the event quarter, though reported up un till three
months later, and the premium volume for the
ompanies parti
ipating
in the pool.
Let us now fo
us on the problems with the ISO index. Let It be the
value of the ISO index at time t. One of the problems was that It was
only published on
e before the settlement date. This took pla
e just
after the end of the reporting period (the Interim report see gure 2).
This meant that the
ompanies, parti
ipating in the pool, had a possi-
bility of knowing at least part of the data used to form the index before
the settlement date, while it was
ertainly more diÆ
ult for other in-
surers. This
reated a information asymmetry whi
h was a potential
fa
tor preventing people from entering the market of CAT-futures.
Another problem was the Moral Hazard problem. A
ompany from
the pool
ould manipulate data by delaying the report of a big loss so it
rst would be in
luded in the next reporting period and thereby never
ae
t the index. The
ompanys intension for doing so,
ould be that
the
ompany had agreed to a short position of a future
ontra
t1 . That
this possibility existed,
ould also have prevented people from entering
the market of CAT-futures.
1 When a
ompany, at time t and at a pri
e Ft , enters a short position of a future
ontra
t, it means that the
ompany should pay (FT Ft ) to the other part of the
ontra
t at time T. The
ompany will then like FT to be as small as possible
6 C. VORM CHRISTENSEN
1 1
Strike value = Comp. loss
omp. market share loss experien
e
At the $4 million atta
hment point:
1 1
Strike value = $ 4 million = $ 2.5 billion or 25 p.
0.2 % 80 %
At the $10 million atta
hment point($ 6 million ex
ess of $ 4 million):
1 1
Strike value = $ 10 million = $ 6.25 billion or 62.5 p.
0.2 % 80 %
Safe-Pla
e's $ 6 million in ex
ess of $ 4 million level of prote
tion is
THE PCS-OPTION, AN IMPROVEMENT 9
Premium
of Call B
A B
PCS LOSS ESTIMATE
Premium
of Spread
If the losses have been estimated below $2.5 billion, the 25/65
all
spread has no value. If the losses have been estimated above $65 billion
Safe-Pla
e re
eives a full prote
tion payment, that is, 40 $200 750,
or $6 million whi
h is the amount of prote
tion originally desired by the
rm. If the losses have been estimated between the 25/65 atta
hment
10 C. VORM CHRISTENSEN
1. Introdu
tion
After Hurri
ane Andrew in 1992 there followed a reinsuran
e
apa
-
ity shortage and a huge in
rease in property
atastrophe reinsuran
e
premiums. Both phenomena where reinfor
ed by the o
urren
e of the
Northridge Earthquake in 1994. The reinsuran
e industry therefore
needed new
apital. This
apital should be found in the nan
ial mar-
ket and the way to obtain it was to
reate the se
uritization market.
One of the rst produ
ts on this market was the CAT future in-
trodu
ed by the CBOT (Chi
ago Board of Trade) in 1992. And ever
sin
e the market has tried to full the investors requirements, but the
market is still not well laun
hed. J. A. Tilley [12℄ is mentioning the
following four reasons why this market is emerging so slowly. First,
sin
e 1994 there has been a generally favourable
atastrophe loss ex-
perien
e and as a result of this the reinsuran
e pri
es have de
reased.
This be
omes a problem be
ause many
edents of risk both primary
writers and reinsurers have
onsidered se
uritization as an alternative
to reinsuran
e rather than
omplementary to reinsuran
e. Se
ond, in-
surers are unwilling to be pioneers, be
ause of the high development
ost. Third, the fa
t that the produ
ts are un
orrelated to other nan-
ial produ
ts is not a good enough selling story for investors. Investors
Date : July 1, 2000.
1991 Mathemati
s Subje
t Classi
ation. 62P05.
Key words and phrases. Catastrophe Insuran
e Derivatives; Derivative pri
ing;
Claims-pro
ess; Heavy tails; Change of measure.
1
2 C. VORM CHRISTENSEN
want to understand the nature of the risk, and this takes time. And
nally there still remains unanswered questions about what form and
stru
ture of insuran
e linked se
urities and derivatives will be viewed
most favourable by investors.
But even though the market have not been well laun
hed and the
produ
ts has not been standardized, a
ademi
s has tried to model the
pri
es of su
h produ
ts see [1℄, [2℄, [5℄, [6℄ and [8℄. And this is also the
aim of this paper. We aim to nd a model that solves two of the main
problems related to pri
ing. The rst problem is that if we
hoose a
realisti
model for the underlying loss pro
ess the market will be in-
omplete and there will exist many equivalent martingale measures.
Hen
e there exists a large set of arbitrage free pri
es of the produ
t.
The next problem is that we like a heavy tail for the underlying loss
index, but heavy tails often give
omputational problems. The model
presented takes in some sense
are of both these problems and is to our
knowledge the rst one to do so. To derive the pri
e of the se
urities
we use results from Gerber and Shiu [10℄. In [10℄ it is shown that the
Ess
her transform is an unique and transparent te
hnique for valuing
derivative se
urities if the logarithms of the underlying pro
ess are gov-
erned by a
ertain sto
hasti
pro
ess with stationary and independent
in
rements (a Levy pro
ess). We propose here su
h a model and by way
of example we
al
ulate pri
es for one the most standardized produ
ts
on the market namely the PCS option.
In the remaining part of this introdu
tion we give a short des
ription
of the PCS option, and the keywords that spe
ify the PCS option are
given. In Se
tion 2 we present the model of the underlying loss index
for the PCS option, and show how the option pri
e is determined within
this model. In Se
tion 3, we show how the results from [10℄
an be used
in our
ontext to nd the risk neutral Ess
her measure. In Se
tion 4
we show how to
ompute the risk neutral Ess
her measure for the loss
period and the development period. In Se
tion 5 we then
al
ulate the
pri
e for the PCS option based on the results obtained in the previous
se
tions. And nally there are some
on
luding remarks.
1.1. Spe
i
ation of the PCS option. In this se
tion the denitions
of the keywords that spe
ify the PCS options are given. For a more
detailed des
ription of the PCS option see [3℄ or [4℄.
The PCS options are traded by Chi
ago Board of Trade and are
regional
ontra
ts whose value is tied to the so
alled PCS Index. The
PCS index tra
ks PCS estimates for insured industry losses resulting
from
atastrophi
events (as identied by PCS) in the area and loss
period
overed. The options are traded as
apped
ontra
ts, i.e. the
ap
limit the amount of losses that
an be in
luded under ea
h
ontra
t.The
value of a PCS
all option at expiration day T , with exer
ise pri
e A
A NEW MODEL FOR PRICING CAT INS DERIVATIVES 3
where N~t is a Poisson pro
ess with a xed parameter 2 and Y~i is
normally distributed (N(; )), where the most natural
hoi
e of
is = 0 (unbiased previous estimates). In order to use the results
from Gerber and Shiu [10℄ we need to assume that the pro
ess Xt for
t 2 [0; T1 ℄ is independent of the pro
ess Xt XT for t 2 [T1 ; T2 ℄. In
1
the real world one will expe
t some dependen
e but the assumption is
invariable in order to use the results form [10℄.
The value of the PCS
all option at expiration day T2 , with exer
ise
pri
e A and
ap value K is given by
C (T2 ; LT ) = min(max(LT
2 2
A; 0); K A):
Let Ft be the information available at time t. We will then make the
following assumption
A NEW MODEL FOR PRICING CAT INS DERIVATIVES 5
where r is the risk free interest rate and E is the mean value a
ording
to a risk neutral measure. Before we
an pro
eed further in the
al
u-
lation of the option pri
e we will have to
hoose a risk neutral measure.
This is done in the next se
tions.
By (3.6) it follows that the
ondition for hl in the loss period is:
exp(r + 1 ) = M (1; 1; hl ) (3.8)
For the development period the situation is similar, the model for
Xt is just dierent. Let now ~2 , ~ and ~ be the adjusted parameters
orresponding to the development period. For t 2 [T1 ; T2 ℄ the premium
is:
Pt = EP [L0 exp(X~ t )℄
= EP [L0 exp(X~ T ) exp(X~ t X~ T )℄
1 1
we obtain
0
V0 = e rt E [t u (mLt =Pt )℄
E [u0 (mLt =Pt )℄
(as a ne
essary and suÆ
ient
ondition, sin
e 00 ( ) < 0 if u00 (x) < 0).
In the parti
ular
ase of a power utility fun
tion with parameter
> 0,
x1
if
6= 1
u(x) = 1
ln(x) if
= 1
We have u0 (x) = x
, and
E [ (m(Lt =Pt ))
℄
rt E [t (Lt =Pt ) ℄
V0 = e rt t = e (3.10)
E [(m(Lt =Pt ))
℄ E [(Lt =Pt )
℄
Formula (3.10) must hold for all derivative se
urities. For t = Lt =Pt
and therefore V0 = 1, (3.10) be
omes
rt E [(Lt =Pt )1
℄
1=e
E [(Lt =Pt )
℄
or
M (1
; t)
exp((r + 1 )t) = : (3.11)
M (
; t)
Comparing (3.11) with (3.7) we see that the value of the parameter
is hl . Hen
e Pt is indeed the expe
tation of the losses Lt
al
ulated
with respe
t to the risk neutral measure.
The results from this se
tion will now be used in the next se
tion,
where the
on
rete risk neutral Ess
her measures for both the loss
period and the development period will be
omputed.
A NEW MODEL FOR PRICING CAT INS DERIVATIVES 9
= exp 2 te h +h e z +(+ h)z 1
2 2
(4.6)
2 2 2
2 2
From (4.5) and (4.6) it follows that the Ess
her transformed pro
ess
is again a pro
ess of the same type as the original one. Let X~t denote
the Ess
her transformed pro
ess then.
N~t
X
X~t = Y~i
i=1
where we now have N~t Po(2 e t) and Y~i N ( 2 h + ; 2).
2 h2 +h
2
2
the solution to (4.7) has to be found numeri
ally.
Finally we will
ompute P (Xt x)
P (Xt x)
= P (X~ t x)
1 n
h +h (2 te hd +hd )n
2
X X 2
P ( Y~i x)
2 2
= exp( 2 te d d )
2
2
n=0
n! i=1
1 2 h 2 +h
X 2 h 2 +h (2 te d d )n x n( 2 hd + )
= exp( 2 te 2 d d )
2
n!
( p 2 )
n=0 n
1
X ( t)n x n
= exp( 2 t) 2 ( p 2 ) (4.8)
n=0
n! n
where 2 = 2 te hd +hd and = 2 hd + . For later use we also
2 2
2
C (t; Lt )
1
X (+1 (T2 t))n
= e( 2 r)(T2 t) L
t exp( 2+1 (T2
t)) 2
n=0
n!
v2 (t) n+1 v1 (t) n+1
( p ) ( p 2 )
n 2 n
1
X ( (T t))n
+e r(T t) exp( 2 (T2 t)) 2 2
2
n=0
n!
v (t) n
K (1 ( 2 p 2 ))
n
v (t) n
A(1 ( 1 p 2 )) (5.2)
n
If t 2 [0; T1 ℄ the value of the
all option at time t is given by
C (t; Lt ) = e r(T t) E [C (T2 ; LT )jFt ℄
2
2
Let now F~ (dx; T1 t; hl ) denote the Ess
her transformed distribution
for X~ (T t) , whi
h we found in se
tion 4.1 (hl denote the parameter
1
whi
h determines the risk neutral Ess
her measure in the loss period).
The value of the option pri
e at time t 2 [0; T1 ℄ is then
C (t; Lt ) = e( 2 r)(T1 t) L0 exp(X )
t
Z 1 1
X (+1 (T2 T1 ))n
exp(x) exp( 2+1 (T2
T1 )) 2
1 n=0
n!
v (t) x n( 2 (h + 1) + )
( 2 p 2d )
n
v1 (t) x n( 2 (hd + 1) + ) ~
( p 2 ) F (dx; (T1 t); hl )
n
Z 1X 1
( (T T1 ))n
+e r(T t) 2
exp( 2 (T2 T1 )) 2 2
1 n=0 n!
v (t) x n
K (1 ( 2 p 2 ))
n
v1 (t) x n ~
A(1 ( p 2 )) F (dx; (T1 t); hl )
n
(5.3)
where F~ (dx; (T1 t); hl ) is given by (4.4)
F~ (dx; (T1 t); hl )
1
X ( (T t))n
= exp( 1 (T1 t)) 1 1 (n; ; x)
n=0
n!
6. Con
lusion
The purpose of this note was to derive a model for pri
ing insur-
an
e derivatives whi
h allows for heavy tails and also provide a unique
pri
ing measure. We su
eeded in nding su
h a model by modelling
the logarithms of the loss pro
ess as a
ompound Poisson pro
ess with
exponential distributed marks in the loss period and with normal dis-
tributed marks in the development period. The pri
e was then found by
evaluating the future payout of the insuran
e derivative under the risk
neutral measure derived by the Ess
her approa
h. We then
al
ulated
the exa
t pri
e in the
ase of the PCS option.
Referen
es
1. Aase, K.K. (1994): \An equilibrium model of
atastrophe insuran
e futures
ontra
ts." Preprint University of Bergen.
2. Barfod, A.M. and D. Lando (1996): \On Derivative Contra
ts on Catastrophe
Losses." Preprint University of Copenhagen
3. Chi
ago Board of Trade (1995): \A User's Guide, PCS options."
4. Christensen, C.V. (1997): \The PCS Option an improvement of the CAT-
future." Manus
ript, University of Aarhus.
5. Christensen, C.V. and S
hmidli, H. (1998): \Pri
ing
atastrophe insuran
e
produ
ts based on a
tually reported
laims." Working Paper Series No. 16,
Centre for Analyti
al Finan
e.
6. Cummins, J.D. and Geman, H. (1993) \An Asian option approa
h to the val-
uation of insuran
e futures
ontra
ts." Review Futures Markets 13, 517-557.
7. Embre
hts, P., C. Kluppelberg and T. Mikos
h (1996): `Modelling Extremal
Events for Insuran
e and Finan
e." Appli
ations of Mathemati
s 33, Springer,
Berlin
8. Embre
hts, P. and S. Meister (1997): \Pri
ing insuran
e derivatives, the
ase of
CAT-futures." In: Se
uritization of Insuran
e Risk: 1995 Bowles Symposium.
SOA Monograph M-FI97-1, p. 15-26.
9. Feller, W. (1971): \An introdu
tion to Probability Theory and its appli
a-
tions." Wiley, Vol. 2. 2nd ed. New York.
10. Gerber, H.U. and E.S.W. Shiu (1996): A
tuarial bridges to dynami
hedging
and option pri
ing. Insuran
e: Mathemati
s and E
onomi
s 18, 183-218.
11. S
hradin, H. R. und Timpel, M. (1996): \Einsatz von Optionen auf
den PCS-s
hadenindex in der Risikosteuerung von Versi
herungsunternehmen
Mannheimer Manuskripte zu Versi
herungsbetriebslehre, Finanzmanagement
und Risikotheorie, Nr. 72.
A NEW MODEL FOR PRICING CAT INS DERIVATIVES 15
1. Introdu
tion
Sin
e the introdu
tion of the insuran
e derivatives in 1992, there have
been a problem pri
ing these produ
ts and several attempts has been
made, see [1℄, [2℄, [5℄, [6℄, [7℄, [9℄ and [12℄. It has not been possible to
nd a unique model like the Bla
k S
holes model be
ause the underly-
ing
annot be des
ribed by a distribution as simple as the log normal
and furthermore, the underlying is not traded. The underlying (the
aggregate
atastrophe losses, whi
h we in the following will denote LT )
would instead most naturally be des
ribed by a marked point pro
ess
with heavy tail distributed marks. But the problem of su
h a model
for the underlying is that the market be
omes in
omplete and it is
then an open question how the pri
ing measure should be determined.
Furthermore, the heavy tailed distribution often gives
omputationally
problems, e.g. if the pri
ing measure is determined by a representa-
tive agent with an exponential utility fun
tion, the marks must have
an exponentially de
reasing tail. So, estimating parameters for the
marked point pro
ess and
al
ulating
onsistent pri
es using a
losed
form pri
ing model is just now not a workable plan.
We therefore lead our analysis in another dire
tion. We follow a pro-
edure familiar to the
onventional option market whi
h also is sug-
gested by Lane and Mov
han in [10℄, namely rather than estimating
volatilities and
al
ulate
onsistent pri
es using, say the Bla
k S
holes
model, take the traded pri
es and extra
t the volatilities
onsistent
with those pri
es, i.e. nd the implied volatility. We
annot use ex-
a
t the same pro
edure on the insuran
e derivative market, sin
e as
1991 Mathemati
s Subje
t Classi
ation. 62P05.
Key words and phrases. Implied loss distribution, parameter estimation, rein-
suran
e, Catastrophe insuran
e derivatives, PCS-options.
1
2 C. VORM CHRISTENSEN
mentioned above, we are not able to
hara
terize the pri
e by a sin-
gle parameter. But we
an do something similar. We
an
hoose a
model for the implied loss distribution and then estimate the implied
parameters from observed pri
es.
This analysis
an be used to evaluate
heapness and dearness among
dierent pri
es and dierent insuran
e derivative produ
ts. We simply
al
ulate implied pri
es from the implied loss distributions and
ompare
them to the observed pri
es. This analysis is very relevant seen in
relation to the re
ent trading su
ess observed at the Chi
ago Board
of Trade (CBoT)
ompetitors namely by The Bermuda Commodities
Ex
hange (BCOE) and The Catastrophe Risk Ex
hange (CATEX).
There are two main problems related to this analysis. First what
kind of distribution should be used for the implied loss distribution
and se
ond, how the involved parameters should be estimated. We are
going to answer these two questions in this paper. The data material
used for this analysis are the pri
es for the National PCS
all spreads
announ
ed by the CBoT on January 1st 1999. For a des
ription of the
PCS-option see [4℄.
The paper will pro
eed as follows. In se
tion 2 we will present the
dierent models for the implied loss distributions, in se
tion 3 we will
des
ribe the pro
edure for estimating the parameters, in se
tion 4 we
present the data and estimate the parameters, in se
tion 5 we evaluate
the dierent models and nally, there are some
on
luding remarks.
The question is now, how is this implied loss distribution of the PCS
index related to the real statisti
al distribution, i.e. the distribution
under the P -measure? Before we try to answer this question we
on-
sider an example.
IMPLIED LOSS DISTRIBUTIONS 3
Example 2.1. Let the statisti
al distribution for the aggregate losses
be des
ribed by a
ompound Poisson pro
ess, i.e.
NT
X
LT = Yi
i=1
where NT Pois(T ), Yi FY and Yi are iid and independent of NT .
We will now pri
e the PCS
all spread by the approa
h of Embre
hts
and Meister [9℄. There the general equilibrium approa
h is used, where
all the utility fun
tions of the agents are of exponential type. More
pre
isely, let at time t the pri
e of the PCS
all spread be given by
PKu;Kl (Lt ; t), the value of the PCS index be given by Lt and the infor-
mation be given by Ft , the pri
e at time t see [9℄, is
E [exp (LT ) min(max(LT Kl ; 0); Ku Kl ) j Ft ℄
PKu;Kl (Lt ; t) = P
EP [exp (LT ) j Ft ℄
where is the risk aversion
oeÆ
ient.
The term exp (LT )=EP [exp (LT ) j Ft ℄ is stri
tly positive and inte-
grates to one. Thus it is the Radon-Nikodym derivative dQ=dP of an
equivalent measure. We
an therefore express the pri
e PKu;Kl (Lt ; t) as
PKu ;Kl (Lt ; t) = EQ [min(max(LT Kl ; 0); Ku Kl ) j Ft ℄
where dQ=dP = exp(LT )=EP [exp(LT ) j Ft ℄. If PKu;Kl (Lt ; t) is the
orre
t pri
e, the distribution of the PCS index under the risk neutral
measure Q should
oin
ide with the implied loss distribution. There-
fore, if we are able to nd the distribution of the PCS index under
the risk neutral measure, we are also able to say something about the
implied loss distribution.
Let us now try to nd this distribution. For an introdu
tion to
hange of measure methods, we refer to [11℄. Let MY () denote the
moment generating fun
tion for Y . By the above denition of the
Q-measure, it follows that 1 i n
Q(NT = n; Yi 2 Ci )
1
= E [exp (LT )1fNT =ng 1fYi 2Ci g ℄
EP [exp (LT )℄ P
n
1 X
= E [exp ( Yi )1fNT =ng 1fYi 2Ci g ℄
EP [exp (LT )℄ P i=1
n
1 Y EP [exp(Yi)1fY 2C g ℄
= MY ()nP (N T = n) i i
exp((MY () 1)) i=1
MY ()
n
1 n ()n Y EP [exp(Y1)1fYi 2Ci g ℄
= MY ( ) e
exp((MY () 1)) n! i=1
MY ( )
n
nY
MY () (MY ()) EP [exp(Y1 )1fYi 2Ci g ℄
= e
n! i=1
MY ( )
4 C. VORM CHRISTENSEN
Hereby it follows that the pro
ess LT is under the new measure Q
a pro
ess of the same type but with dierent parameters as under P .
Under Q, NT is a Poisson pro
ess with rate MY (R) and the individual
laims have the distribution fun
tion FYQ(x) = 0x ey dF (y )=MY ()
(e.g. if Y (;
) then Y ( ;
) under Q). Similar results for
another model
an be found in [5℄.
In example 2.1 we nd that the implied loss distribution and the sta-
tisti
al distribution are of the same type only with dierent parameters.
It would be
onvenient if this were true in general. When modelling
the implied losses, we would only have to look among the models whi
h
reasonably
ould be used to des
ribe the real losses. But is this true in
general?
As mentioned earlier the most natural way to des
ribe the real losses
is by a marked point pro
ess with positive marks. Let us now re
all the
denition of a marked point pro
ess. A marked point pro
ess with posi-
tive marks is a sequen
e (Tn ; Yn)n1 of sto
hasti
pairs, where T1 ; T2 ; : : :
are non-negative and represent time of o
urren
e of some phenomena
represented by the non-negative elements Y1 ; Y2; : : : referred to as the
marks of the pro
ess. By this denition it follows that if (Tn ; Yn) is
a marked point pro
ess with positive marks under P , then (Tn ; Yn) is
a marked point pro
ess with positive marks under Q, where Q is an
equivalent measure. So, if the real losses are des
ribed by a marked
point pro
ess the implied losses should also be des
ribed by a marked
point pro
ess.
The distribution for the implied losses and the distribution for the
real losses will in general not be the same. But by the dis
ussion above,
we will only use models whi
h
ould reasonably be used to des
ribe the
real losses, when we now start to model the implied losses. We now
present six models for the implied losses.
2.1. Model 1. The rst model we will use in our analysis is the same
model as the one suggested by Lane and Mov
han [10℄, namely a
om-
pound Poisson model with gamma distributed
laims, i.e.
NT
X
LeT = Yi
i=1
tailed, whereas data give eviden
e that the distribution tail of the ag-
gregate
laims is heavy tailed. In this model we
an only approximate
a heavy tail by
hoosing low values of
and .
It is important to have this model in the analysis in order to see
how the result from this model diers from the following and more
ompli
ated models. The value of the PCS
all spread with strikes Kl
and Ku at time 0 is given by
PKu;Kl (L0 ; 0)
Z Ku
= (x Kl )feLT (x)dx + (Ku Kl )(1 FeLT (Ku ))
Kl
1 n Z Ku
X
= e n
(n
) 1 xn
e x dx
n=1
n! Kl
Z 1
+ Ku n
(n
) 1 xn
1 e x dx
K
Z 1u
Kl n
(n
) 1 xn
1 e x dx
Kl
where L0 = 0.
2.2. Model 2. Looking at the listed
all spreads we see, that the one
with the lowest strikes is the 40/60
all spread. The bid and ask for this
all spread is 12 and 15 respe
tively, whi
h are relatively large values
for a produ
t that has a maximal pay-out of 20. These fa
ts
ould
therefore indi
ate that the market expe
ts that the loss index will be
above a given threshold K0 for sure. If this is true and K0 > 40,
then there is no market for a 20/40
all spread, be
ause the market
will expe
t the
all spread to be worth 20 for sure. Based on these
indi
ations we extend model 1,
NT
X
LeT = K0 + Yi
i=1
where LeT are the implied losses, K0 is a
onstant indi
ating the thresh-
old the market expe
ts the losses to be above for sure, NT Pois(T )
and Yi (
; ). A
ording to the bid of the 40/60
all spread (12),
we will not allow K0 to be above 52 (40+12). A possible interpretation
of this model is to think of K0 as the mean value of the \normal"
laims
and of the
ompound Poisson pro
ess as a model of the ex
esses.
6 C. VORM CHRISTENSEN
The
omputations are still very simple. The value of the PCS
all
spread with strikes Kl and Ku at time 0 is here given by
PKu;Kl (L0 ; 0)
Z Ku
= (x Kl )feLT (x)dx + (Ku Kl )(1 FeLT (Ku ))
Kl
1 n Z Ku K0
X
= e n
(n
) 1 xn
e x dx
n=1
n! (Kl K0 ) +
Z 1
+ (Ku K0 ) n
(n
) 1 xn
1 e x dx
K K
Z 1u 0
(Kl K0 ) n
(n
) 1 xn
1 e x dx
(Kl K0 )+
re
all that we require K0 < 52 so the term (K0 Kl )+ is only in
luded
in the pri
e of the 40/60
all spread.
2.3. Model 3. The next model will also rely on a light tail distribution
but we will now put more
u
tuation into the model.
The PCS index
an be viewed as the sum of losses from the individual
atastrophes, and the losses from the individual
atastrophe
an be
viewed as the sum of the individual
laims
orresponding to this single
atastrophe. We
ould therefore model the PCS index LT as
NT X
X Mi
LT = Yij
i=1 j =1
where NT is the number of
atastrophes, Mi the number of
laims from
the ith
atastrophe and Yij is
laim size number j from
atastrophe
number i. A similar model is also suggested in [6℄, here a
losed form
pri
ing model are derived and it is shown how the above model
an be
used to in
orporate the reporting times of the
laims.
The number of
laims from a
atastrophe is P very large, so by the
strong law of the large numbers it follows that M j =1 Yij Mi E [Yij ℄.
i
-
20 40 60 80 100 120 Strike
3. From this plot we
hoose a fun
tion to des
ribe , i.e if we
hoose
a fun
tion f with three parameters a, b and
, a; b;
2 R. We
an now des
ribe by (K ) = f (a; b;
; K ). The implied loss
distribution for a PCS
all option with strike K is therefore given
by
LeT Pa((K ); )
where (K ) = f (a; b;
; K ).
4. Let now PK (0; L0 ) Rdenote the value of a PCS
all option at time
0, i.e. PK (0; L0 ) = K1(x K )feL (x)dx. The value of the PCS
all
IMPLIED LOSS DISTRIBUTIONS 9
su
h that they generate pri
es that are (i) lower than known oers; (ii)
higher than known bids, and (iii)
losest to a
tual traded pri
es. The
optimization is two-tier. First, get inside the bid-oer spread. Se
ond,
get
losest to a
tual traded pri
es. The two-tier ee
t is a
hieved by at-
ta
hing (ideally non-Ar
himedean) weights to ea
h of the two obje
tive
fun
tion. \Closest" is dened as the absolute value of the dieren
e
between the a
tual traded pri
e and the theoreti
al (or tted) pri
es".
We agree that it is desirable that the parameters are
hosen su
h that
the pri
es fulll (i) and (ii), but we do not think that the requirements
should be invariable be
ause, if the spreads are very small, it
ould be
a problem to nd a solution. And if the theoreti
al pri
es appear to be
far away from the spread, it
ould be used to indi
ate that the
hosen
model may be wrong. We also agree on point (iii), i.e. if our data
ontain only traded pri
es, the parameters should be found by a least
square t. But the data primarily
onsist of spreads and single bids or
asks, we therefore suggest the following obje
tive fun
tion.
3.1. The obje
tive fun
tion. The obje
tive fun
tion O that we pro-
pose be minimized in order to nd the parameters is the following
X P bid P th + 2 X P th P ask + 2
O = i i + i i
P bid P ask
bids i asks i
| {z } | {z }
term 1 term 2
1 X (Pi bid ask
Pi )
+ Æ1
#spreads spreads (Pibid + Piask)=2
| {z }
term 3
X P th (P bid + P ask )=2 2 1
i
Pi
i
bid Pi ask
i ^4
spreads
| {z }
term4
X P th 2P bid + 2
+ Æ2 i i
P bid
single bids i
| {z }
term 5
X 0:5P ask P th + 2
+ Æ2 i i
P ask
single asks i
| {z }
term 6
where the Pith 's are the theoreti
al pri
es, the Pibid's are the observed
bids and the P ask 's are the observed asks. Æ1 and Æ2 are both
onstants.
i
When Pi is a traded pri
e, Pi is
onsidered as both a bid and an ask
where Pibid = Piask.
Term 1 and 2 are in
luded be
ause as mentioned above we prefer the
theoreti
al pri
es to be above the observed bids and below the observed
asks. And as for the optimal
ase where we have only traded pri
es and
IMPLIED LOSS DISTRIBUTIONS 11
no bid/ask spreads, these two terms alone will give us the
ommonly
used least square t.
As long as the average length of the spreads is small we are
lose to
the optimal
ase where we only have traded pri
es and term 1 and 2
will probably be suÆ
ient to nd a solution. But if the average length
of the spread is large there is less information about the pri
es and we
will probably be unable to nd a unique solution. We therefore add
term 4, in this term we value the information from the bid and the ask
equally, i.e. we prefer the theoreti
al pri
e to be in the middle of the
bid/ask spread. We
ap the single terms in the sum at 1/4, be
ause if
Pith = Pibid or Pith = Piask the single term in the sum is equal to 1/4,
and if Pith > Pibid or Pith < Piask then it is punished in term 1 or 2. How
mu
h this fourth term should be valued
ompared to term 1 and 2 is
then adjusted by term 3. Term 3 is a
onstant Æ1 and a term denoting
the average length of the spread. In agreement with the
omments
above, we thereby obtain, that if the average length of the spreads is
small, we weight Pith being in the middle less than if the average length
of the spreads is large.
The term 5 and 6 are in
luded in order to se
ure that the theoreti
al
pri
es do not get too far away from the single bid's or ask's. By too far
away we mean that a theoreti
al pri
e is punished if it is lower than
50 % of a single ask or higher than 200 % of a single bid. By the term
Æ2 we are able to adjust how mu
h the fth term should be valued
ompared to the other terms.
If we also had information about the volumes that are bid and asked,
it
ould be argued that the bid and ask pri
es should be weighted by
the
orresponding volumes. This is due to the fa
t that if the volume
is large, the traders are more
on
erned about the pri
e, and therefore
the pri
e should be more a
urate. The problem with this argument
is that there is an opposite ee
t, namely if the asks are mu
h higher
than the \true" pri
e, we also expe
t the volumes to be large and if
the bids are mu
h lower than the \true" pri
e, we expe
t the volume
to be large. We therefore do not take the volumes into the obje
tive
fun
tion. We are now ready to estimate the parameters for the models
des
ribed in se
tion 2 by the above obje
tive fun
tion. This estimation
is the subje
t of the next se
tion.
The rst
hange in the underlying PCS index was made January
19th, where the index in
reased from 0 to 7.6. We have
hosen the
data from January 7th be
ause the last
hanges in the bids and asks
before January 19th were made here. If we take data from dates after
January 19th we have to take the value of the index into a
ount. If we
onsider data from a time point t where the PCS index is greater than
0, some adjustments have to be done. The implied losses at expiration
time T
an, at time t, be written as LeT = (LeT Lt ) + Lt , where Lt
is a
onstant and LeT Lt is the implied losses in the period from t
to T . LeT Lt
an then be des
ribed by the same models as we used
to des
ribe LeT , but the parameters will probably be
hanged. Even
though we are looking at a model where LeT is a stationary pro
ess, we
annot expe
t the same parameters sin
e the PCS index is in
uen
ed
by some large seasonal ee
ts.
The National PCS
all spreads announ
ed by the CBoT on January
7th 1999 is given by table 4.1.
are very brief. But from the table it is very
lear that model 3 and 4
seems to be
omputationally heavy
ompared to model 1, 2, 5 and 6.
We will keep this in mind for the evaluation of the models. It is not
possible to disqualify any of the models, be
ause the
omputer times
ould be redu
ed with a faster
omputer, or by using method tailored
for the problem to solve.
4.3. The theoreti
al pri
es. We are now ready to estimate the pa-
rameters. The estimation of the parameters for model 1, 2, 3, 4 and 6 is
straightforward. We simply write a
omputer program that
al
ulates
the theoreti
al pri
es a
ording to the above formulas and minimize
the obje
tive fun
tion a
ording to the des
ription in [3℄ and [8℄. For
model 5 we have to do the estimation in four steps as des
ribed in
se
tion 2. In step 1 we nd = 7:66 and = 508. In step 2 we then
set = 508 and estimate the dierent 's, the 's are shown in gure
4.1.
6
-
50 100 150 200 250 300 Strike
The parameters found by minimizing the obje
tive fun
tion, the
or-
responding mean values and varian
es for the implied losses and the
theoreti
al pri
es are listed in table 4.3, table 4.4 and table 4.5 respe
-
tively.
M1 M2 M3 M4 M5 M6
Mean value 74 90 77 96 (73;91) 139
Varian
e 6096 8453 6178 11652 1 1
Table 4.4. Mean value and varian
e.
was su
essfully suggested by Lane and Mov
han in [10℄ so why now
this dieren
e? In [10℄ they
onsider market pri
es midyear 1998 where
the PCS index was nearly 40 and this apparently makes a dieren
e.
We also tried to model the midyear 1998 pri
es with model 1 and our
obje
tive fun
tion. The results are shown in table 5.1 (The parameter
from [10℄ has been adjusted to
orrespond to the index value and not
the Billion $ value).
Kl =KU bid LM CVC ask
40/60 11.0 11.0 12.0
60/80 6.0 7.5 8.2 10.0
80/100 5.7 6.1 8.0
100/120 3.5 4.4 4.6 6.0
100/150 9.4 9.5 12.0
120/140 1.0 3.5 3.5 6.0
250/300 0.5 1.9 1.4 2.5
100/200 14.7 14.4 20.0
150/200 4.0 5.4 4.9 7.5
180/200 0.4 1.8 1.6 1.8
2.23 2.17
0.1887 0.2645
0.0089 0.0124
Table 5.1. The data from [10℄
ontra our data.
From table 5.1 we see that model 1 in our obje
tive fun
tion also
generates reasonable results for the midyear 1998 pri
es. We therefore
on
lude that the reason for the bad t of the 1999 pri
es is the model
and not the obje
tive fun
tion. Another important thing to note from
table 5.1 is that there are remarkable dieren
es in the pri
es obtained
by Lane and Mov
han and the pri
es we obtain. We thereby see that
the valuation of the bid's and ask's is highly dependent on the
hoi
e
of the obje
tive fun
tion.
5.2. Model 2. Looking at the results from model 2 we see a remark-
ably better t. All the pri
es are now in the bid/ask spreads and we
also have reasonable pri
es for the 250/300 and 300/350
all spreads.
So the shift of the distribution to x = 47:2 apparently has a large ee
t.
This is in agreement with the results from [10℄. If we
ompare the 1999
pri
es with the midyear 1998 pri
es we see that they are quite similar,
see table 5.2
Looking at the gures in table 5.2 and re
alling that the
ompound
Poisson distribution ts the midyear 1998 pri
es (where the index value
was 40), it is not suprising that the shifted
ompound Poisson distribu-
tion ts the 1999 pri
es. Model 1 is a spe
ial
ase of model 2 (x = 0),
so model 2 will also be able to t the midyear 1998 pri
es.
16 C. VORM CHRISTENSEN
5.3. Model 3. From table 4.5 we see that model 3 generates pri
es
very similar to model 1, so we do not obtain mu
h by in
luding the
extra parameter. And as for model 1 we
on
lude that model 3 is a
bad des
ription of the implied losses. As done for model 1 we
ould
also extend model 3 by shifting it, but we will desist from doing this
be
ause we will then have ve parameters, whi
h we
onsider too many.
How many parameters one will allow in a model is of
ourse individual,
but we set the limit by four. We dis
uss this further in the di
ussion
of model 5.
5.4. Model 4. A very important thing to note about this model is
that the
omputation of the
onvolutions is very time
onsuming and
this is also why we only in
lude the rst four
onvolutions in the model.
The pri
es we obtain by in
luding only the rst four
onvolutions are
very similar to the pri
es from model 1 and model 3, and as for model
1 and model 3 it is not possible to get into the 40/60 and 200/250
all
spreads. The pri
es for the 250/300 and the 300/350
all spreads seem a
little bit better than those forPmodel 1 and 3. The model was justied
by assuming that the term 1 n
5 e =n! should be small. Here the
term is 0.12, whi
h
an hardly be
onsidered small. So, with only four
onvolutions, we
onsider the model as a bad des
ribtion of the implied
losses.
If we had in
luded more
onvolutions, 10 say, we would have ex-
pe
ted a remarkably better t, but this would have been way to time-
omsuming. The time fa
tor is also the reason why we desist from
shifting the distribution.
5.5. Model 5. The rst thing to note from model 5 is the number
of data being very small, whi
h therefore makes it hard to really gain
anything from gure 4.1 (the relation between the -parameter and
the strike). The fun
tion we
hoose based on gure 4.1 is therefore
also very general. The three parameters in
luded in the fun
tion are
many
ompared to the number of data. It is therefore not surprising
that we get a better t than for model 1, 3 and 4, and if we had
in
reased the number of parameters even more, say to four or ve, we
would probably also obtain a better t than we did in model 2. But
IMPLIED LOSS DISTRIBUTIONS 17
in
reasing the number of parameters does not make the model any
better in relation of des
ribing the implied losses. After the estimation
of the pri
es for model 5 it is our general impression that model 5 is a
bad des
ription of the implied losses. We thereby do not
on
lude that
there is no strike dependen
y on the inplied parameters, but if there
is, it has to be modeled in another way.
5.6. Model 6. It is very fast to
ompute the pri
es
orresponding to
model 6 and the t we obtain is surprisingly good. Suprisingly be
ause
it is not a
ompound pro
ess, but only a shifted Pareto distribution,
and it generates pri
es that are better than the pri
es from model 2.
But even though the model generates a good t, the pri
es are very
dierent from the pri
es obtained by model 2. In general it derives
pri
es that are higher than the pri
es from model 2. The dieren
es
in pri
es are not surprising if we
ompare the mean values and the
varian
es from model 1 with the ones from model 6, see table 4.4.
We now have two models both generating reasonable pri
es and hav-
ing nearly the same value of the obje
tive fun
tion. But this is not the
same as saying that the two models are equally good. Based on the
dis
ussion in se
tion 2, model 2 seems to be the best theoreti
ally
founded model. But if we look at the model 2 pri
es for the 60/80 and
the 80/100
all spread, one
ould get the impression that the implied
distribution from model 2 generates too little risk, whi
h indi
ates that
model 2 is not a perfe
t model.
After this dis
ussion we nd that model 2 is the best model to use
even though it is not perfe
t. But we also nd that one should use
model 6 simultaneously be
ause it is very fast and it
ould be used to
support the evaluation of the pri
es.
5.7. The Æ1 and Æ2 parameter. How the parameters Æ1 and Æ2 should
be set depends on the data set. If the data set
onsist of only bid/ask
spreads and our model generates pri
es that are inside these spreads,
then the value of Æ1 has no ee
t on the results. But if we have also
traded pri
es in our data set, the value of Æ1 gets more important. In
this
ase we believe that one should estimate pri
es for dierent values
of Æ1 and then set the parameter based on an evaluation of these results.
Model 2 and model 6 both generate results that are inside the bid/ask
spread and the value of Æ1 therefore be
omes unimportant for the data
set
onsidered in this paper.
The value of Æ2 is set to 0.1, i.e. if the pri
es are more than 100%
above a single bid or 50% below a single ask we punish it only 10%
as hard as when the pri
es are below or above a bid or an ask. We
nd this value reasonable but again we
ould for a given data set try
with dierent values and based on these estimations set the parameter.
But again for model 2 and model 6, the value of Æ2 does not ae
t the
results in this paper.
18 C. VORM CHRISTENSEN
6. Con
lusion
We analyse pri
es for
atastrophe insuran
e derivatives by looking
at the \implied loss distributions" embedded in the traded pri
es. As
mentioned in the introdu
tion there are two main problems in this anal-
ysis. First, what kind of distribution should be
hosen for the implied
losses and se
ond, how should the involved parameters be estimated?
In relation to how the parameters should be estimated we nd that
an improvement of the pro
edure from [10℄ was ne
essary be
ause of
the following two reasons. First we agree that it is desirable that the
parameters are
hosen su
h that the pri
es are lower than known oers
and higher than known bids, but we do not think that the requirements
should be invariable be
ause, if the spreads are very small, it
ould be
a problem to nd a solution. And if the theoreti
al pri
es appear to be
far away from the spread, it
ould be used to indi
ate that the
hosen
model may be wrong. Se
ond we agree on point that the parameters
should be
hosen su
h that the pri
es gets
losest to the a
tual traded
pri
es, i.e. if our data
ontain only traded pri
es, the parameters should
be found by a least square t. But be
ause the data primarily
onsists
of spreads and single bids or asks, we nd that this should be in
or-
porated in the obje
tive fun
tion. No matter what obje
tive fun
tion
one uses, it is
lear from the dis
ussion of model 1 that the
hoi
e of
the obje
tive fun
tion has a large ee
t on the derived pri
es, and it
should therefore be
hosen
arefully.
After the dis
ussion in the previous se
tion we nd that model 1,
suggested by Lane and Mov
han [10℄, is unable to t the PCS-option
pri
es in general. Instead, we nd that model 2 is a better model to use
for the implied losses. However, it would be preferable to use model
6 also in order to support it. It is
lear that none of the suggested
models t the implied losses perfe
tly, but we believe that model 2
supported by model 6 will be a good tool for investors analysing pri
es
of
atastrophe insuran
e derivatives.
A
knowledgment
The author thanks Uwe for a fruitful dis
ussion on the topi
.
Referen
es
1. Aase, K.K. (1994), An equilibrium model of
atastrophe insuran
e futures
on-
tra
ts, Preprint University of Bergen.
2. Barfod, A.M. and D. Lando (1996), On Derivative Contra
ts on Catastrophe
Losses, Preprint University of Copenhagen.
3. Broyden, C.G. (1970), The
onvergen
e
lass of double-rank minimization al-
gorithms, J. Inst. Math. Appl.
4. Chi
ago Board of Trade (1995), A User's Guide, PCS options.
5. Christensen, C.V. (1999), A new model for pri
ing
atastrophe insuran
e deriva-
tives, Working Paper Series No. 28, Centre for Analyti
al Finan
e.
IMPLIED LOSS DISTRIBUTIONS 19
6. Christensen, C.V. and S
hmidli, H. (1998), Pri
ing
atastrophe insuran
e prod-
u
ts based on a
tually reported
laims, Working Paper Series No. 16, Centre for
Analyti
al Finan
e.
7. Cummins, J.D. and Geman, H. (1993), An Asian option approa
h to the valu-
ation of insuran
e futures
ontra
ts, Review Futures Markets 13, 517-557.
8. Fielding, K. (1970), Fun
tion minimization and linear sear
h, Communi
ation
of the ACM, vol 13 8, 509-510.
9. Embre
hts, P. and S. Meister (1997): Pri
ing insuran
e derivatives, the
ase
of CAT-futures. In: Se
uritization of Insuran
e Risk: 1995 Bowles Symposium.
SOA Monograph M-FI97-1, p. 15-26.
10. Lane, M. and Mov
han, O. (1998), The perfume of the premium II, Sedwi
k
Lane Finan
ial, Trade Notes.
11. Rolski, T., H. S
hmidli, V. S
hmidt and J.L. Teugels (1999), Sto
hasti
pro-
esses for insuran
e and nan
e. Wiley, Chi
hester.
12. S
hmo
k, U. (1998), Estimating the value of the WINCAT
oupons of the Win-
terthur insuran
e
onvertible bond: A study of the model risk, ASTIN bulletin
29.
13. Tilley, J.A. (1997), The se
uritization of
atastrophe property risks, Pro
eed-
ings, XXVIIth International ASTIN Colloqium, Cairns, Australia.
1. Introdu
tion
During the last years, the market for risk related to natural phenom-
ena su
h as dierent
atastrophes has witnessed important
hanges.
Su
h risk have traditionally been distributed through the insuran
e
and reinsuran
e system. Insuran
e
ompanies a
umulate the risk of
individual entities and redistribute the risk to the global reinsuran
e
industry. But the volatility of weather, taken together with population
movement to warm
oastal areas and
hange of property pri
es has
made
atastrophi
risk highly unpredi
table. It is therefore no longer
possible to diversify this risk using traditional insuran
e pra
ti
es. A
new way to manage su
h risk or unknown risk in general is
alled for.
When we talk about unknown risk, we refer to risk whi
h frequen
y
we do not know, i.e. there is more than one estimate of the frequen
y
of the risk. Examples of unknown risk are environmental health risk of
new and little known epidemi
s, or risk indu
ed by s
ienti
un
ertainty
in predi
ting the frequen
y and severity of
atastrophi
events.
As we will show in this paper, motivated by [2℄, the way to han-
dle unknown risk is to use two dierent approa
hes of hedging risk
simultaneously, namely the statisti
al approa
h known from the insur-
an
e industry and the e
onomi
approa
h known from the se
urities
industry.
In this paper we
onsider a general model for an insuran
e
ompany,
where the
ompany fa
es n states of the world. For ea
h of these states
Date : June 30, 1999.
1991 Mathemati
s Subje
t Classi
ation. 91B99.
Key words and phrases. Unknown risk, interplay between insuran
e and nan
e,
atastrophe insuran
e,
atastrophe insuran
e derivatives.
1
2 C. VORM CHRISTENSEN
P = q1 P1 + + qn Pn
E [O1 ℄ E [O ℄
p =p i ; i = 2; : : : ; n
Var(O1 ) Var(Oi)
P0 = P1 q1 + + Pn qn
Note that we here have V ar(Oi) = V ar(Li ). OPC1 therefore ensures
that expe
ted gain E [Pi Li ℄ will be high in states where V ar(Li ) is
large, i.e. we obtain a high gain in the risky states.
4.2. OPC2: Equal ruin probabilities in all states. The goal here
is to obtain the same ruin probabilities in all the states. The ruin
probabilities are usually
al
ulated a
ording to the initial
apital, we
therefore in
lude the initial
apital u in the OPC. The solution is found
by solving the following n equations with n unknowns:
E [v (u + O1 )℄ = E [v (u + Oi )℄; i = 2; : : : ; n
P0 = P1 q1 + + Pnqn
4.4. OPC4: Maximal expe
ted utility. The goal here is to obtain
the maximal expe
ted utility. We have the same utility fun
tion as in
OPC3, but we now have to solve the following maximization problem:
max E [v (u + Ps Ls )℄
st P0 = P1 q1 + + Pn qn
We have now stated four dierent ways of solving the OPC, but
whi
h one is the best? At rst it seems most natural to use OPC4,
HOW TO HEDGE UNKNOWN RISK 7
i.e. maximize the expe
ted utility. But a problem by this approa
h
is that when we only
onsider the general expe
ted utility we
ould
end up with some very risky individual states. One
ould imagine a
situation where the insuran
e
ompany maximizes its expe
ted utility
and thereby obtains a very large ruin probability in one of the states.
Su
h a premium
hoi
e
ould then, be
ause of the high ruin probability,
be refused by shareholders, authorities or the like. OPC4 is therefore
not ne
essarily the best OPC to use in general. In the following we will
analyse the OPC's further.
5.1. Solution of OPC1. Let now i denote the p mean value and i
denote the standard deviation of Li . We then have Var(Oi ) = i and
OPC1
an now be written as
P1 1 Pi i
= ; i = 2; : : : ; n
1 i
P0 = P1 q1 + + Pnqn
The general solution to these n equations with n unknowns is given by
Pi = ai P1 + bi ; i = 2; : : : ; n
P q b qn bn
P1 = 0 2 2
q1 + q2 a2 + + qn an
P 00 (r) v 00 (x)
q2 10 2 = 0 :
P1 (r) v (x)
The right hand side is independent of r and the left hand side is inde-
pendent of x. The left hand side and the right side is therefore both
independent of r and x. And from this we
on
lude
v 00 (x)
=
v 0 (x)
for some 0 and the assertion follows.
Example 5.9. If we now set v (x) = A(1 e x ) we are able to nd
a
losed solution to OPC3. Let MLi () denote the moment generating
fun
tion of Li at the point . From (5.3) we then know that OPC3
an be written as
1 ML ( )
Pi = P1 ln(
1
); i = 2; : : : ; n
MLi ()
P0 = P1 q1 + + Pn qn
We now have a problem equivalent to OPC1 and the general solution
is therefore given by
1
P1 = P0 + (q2 ln(a2 ) + + qn ln(an ))
1
Pi = P1 ln(ai ); i = 2; : : : ; n
L ()
where ai = M MLi () .
1
HOW TO HEDGE UNKNOWN RISK 11
max p1 E [v (u + P1 L1 )℄ + + pn E [v (u + Pn Ln )℄
st P0 = P1 q1 + + Pn qn
In order to solve this problem we
onstru
t the Lagrange fun
tion
F (P1 ; : : : ; Pn) = p1 E [v (u + P1 L1 )℄ + + pn E [v (u + Pn Ln )℄
+(P0 P1 q1 Pn qn )
The next step is to
onstru
t the rst and se
ond order
ondition
or-
responding to the maximization problem. The rst order
ondition:
F
= p1 E [v 0 (u + P1 L1 )℄ q1 = 0
P1
..
.
F
= pn E [v 0 (u + Pn Ln )℄ qn = 0
Pn
F
= P0 P1 q1 Pn qn = 0
Let now gi(x) = E [v 0 (u + x Li )℄, then gi (x) is stri
tly de
reasing
and
ontinuous, hen
e the inverse exists, whi
h also is
ontinuous and
de
reasing. From equation number 1 and i, i = 2; : : : ; n we get
pq
Pi = gi 1( 1 i g1 (P1 )): (5.11)
pi q1
P1 is found by plugging (5.11) into the
onstraint, i.e. from the equation
pq pq
P0 = q1 P1 + q2 g2 1 ( 1 2 g1 (P1 )) + + qn gn 1 ( 1 n g1 (P1 )):
p2 q 1 pn q 1
1 p q
Note that q1 P1 and qi gi ( pi q i g1 (P1 )) for all i are stri
tly in
reasing in
1
1
P1 . It is therefore possible to nd a solution to this equation.
We now have a solution. In order to
he
k whether it is optimal we
he
k the se
ond order
ondition. To
he
k the se
ond order
ondition,
for a minimization problem with an equality
onstraint, we have to
onstru
t the so
alled bordered determinants, see [3℄ (p. 382). They
are obtained by bordering the prin
ipal minors of the Hessian deter-
minant of se
ond partial derivatives of the Lagrange fun
tion by a row
and a
olumn
ontaining the rst partial derivatives of the
onstraint.
The element in the southeast
orner of ea
h of these arrays is zero.
12 C. VORM CHRISTENSEN
The se
ond order
ondition for the problem is then satised if these
bordered determinants alternate in sign, starting with plus, i.e. the
sign of the determinants below must taken from above be +,{,+, et
.
2 3
p1 E [v 00 (u + P1 L1 )℄ 0 q1
4 0 00
p2 E [v (u + P2 L2 )℄ q2 5
q1 q2 0
..
. 2 3
p1 E [v 00 (u + P1 L1 )℄ : : : 0 q1
6 .. ... .. .. 7
6 . . . 7
4 0 : : : pn E [v 00 (u + Pn Ln )℄ qn 5
q1 ::: qn 0
It follows easily from the bordering determinants above that the se
ond
order
ondition is fullled.
In the following example we solve the problem expli
itly for the ex-
ponential utility fun
tion.
Example 5.12. As for OPC3 we now solve OPC4 expli
itly for v (x) =
A(1 e x ). The OPC4 problem then takes the following form:
max 1 p1 E [e (P L ) ℄ pn E [e (Pn
1 1 Ln ) ℄
st P0 = P1 q1 + + Pn qn
As before we rst
onstru
t the Lagrange fun
tion
F (P1 ; : : : ; Pn ) = 1 p1 ML ()e 1
P1 pnMLn ()e Pn
+(P0 P1 q1 Pnqn )
The next step is then to
onstru
t the rst and se
ond order
ondition
orresponding to the maximization problem. The rst order
ondition
as before:
F
= p1 ML ()e P q1 = 0 1
P1 1
..
.
F
= pn MLn ()e Pn qn = 0
Pn
F
= P0 P1 q1 Pn qn = 0
From equation number 1 and i, i = 2; : : : ; n we get
1 p q M ( )
Pi = P1 ln ai ; where ai = 1 i L : (5.13) 1
pi q1 MLi ()
HOW TO HEDGE UNKNOWN RISK 13
Pi = E [Li ℄ + P0 EQ [L℄
= E [Li ℄ + 0:975
This is also what we expe
t, an agent with a risk neutral behaviour
allo
ates the same amount of safety loading in all the states.
Finally
onsider OPC4. When we
onsider OPC4 we have to remem-
ber that OPC4 both in
lude the p and the q probabilities. This means
that the premium
hoi
es for an agent using OPC4 be
omes depended
on how he evaluates the market pri
es, i.e. how his preferen
es are
ompared to the preferen
es of the representative agent. This means
that if the agent
onsider the pri
e of AD se
urity i (qi ) to low
om-
pared to his own preferen
es he will be buying more of AD se
urity i.
But when he buys more of AD se
urity i his portfolio
hanges and at
some level he will
onsider the pri
e to high. So when we now evaluate
OPC4 we will have to remember that the agent maximize his utility
and evaluate the pri
es (the q probabilities) at the same time.
For a high risk aversion
oeÆ
ient we see, as for OPC3, that the
agent sells out of AD se
urity 1 and 2 in order to buy more of AD
16 C. VORM CHRISTENSEN
se
urity 3. We also see that for = 0:09 P3 is lower for OPC4 than
for OPC3. The reason for this is as mentioned above that when the
agent using OPC4 obtain a premium P3 = 44:543 in state 3 he starts
onsidering the pri
e of AD se
urity 3 (q3 ) to high.
The values of = 0:04549 and = 0:04384 are in
luded in the table
in order to see how OPC 4 is related to OPC1 and OPC2. For the
two values we obtain P1OP C 3 = P1OP C 1 and P1OP C 3 = P1OP C 2. We here
observe that OPC4 put more weight into state 3 that both OPC1 and
OPC2.
Finally we see that for ! 0, i.e. we are moving towards a risk
neutral behaviour the agent, as expe
ted,
onsiders the pri
e of AD
se
urity 3 higher and higher. He therefore sets P3 lower and lower, i.e.
selling more and more of AD se
urity 3.
6.2. Example 2. As mentioned above we here assume that the total
losses in state 1 is exponentially distributed (L1 Exp(1 )) and the
total losses in state 2 is Pareto distributed (L2 Pa(2 ; 2 )). The
possible premium (P0 ) is here set to 1:2 EQ [L℄ and the q probabilities
are set to q1 = 0:75 and q2 = 0:25. In this example we
onsider three
dierent values for the parameters. This is done in order to see how
the two OPC's (OPC1 and OPC2) dier when the loss distribution in
state 2 be
omes more and more heavy tailed. The parameter values,
the mean values and the varian
es for the Li 's, the OPC, the risk
quantity's and the ruin probabilities are all given in the table below.
OPC2 is solved for dierent values of the initial
apital. The value of
the u's are given in the table.
If we rst
onsider OPC1, we see that the value of P2 in
reases and
the risk quantity de
reases when L2 be
omes more and more heavy
tailed. So again OPC1 works the way we want, i.e. we have a relatively
large safety loading in the states where our risk is high.
We now
onsider OPC2. Here the pi
ture is dierent. For u = 0
and u = 10 we surprisingly observe that the the value of P2 and the
ruin probability both de
reases when L2 be
omes more and more heavy
tailed. But for u = 20 we observe a more preferable behaviour, namely
that P2 and the ruin probability both in
reases when L2 be
omes more
and more heavy tailed. The reason for the \bad" OPC for u = 0 and
u = 10 must be found in the shape of the distribution fun
tions. In
gure 1 and gure 2 P (L < P + u) (1-"the ruin probability") are shown
for u = 0 and u = 20 respe
tively. The labels
orresponds to the fol-
lowing distributions of L Exp(0:1)(0), Pa(6; 50)(1) and Pa(2:1; 11)(3).
The horizontal lines are the levels of the survival probability (1-ruin
probability) for OPC2.
If we start analysing gure 1, it be
omes
lear why we observe the
de
reasing P2 when L2 be
omes heavier tailed. Be
ause of the low val-
ues of the initial
apital we have a relative low level of the distribution
HOW TO HEDGE UNKNOWN RISK 17
0.8
0.6
0.4
0.2
0 20 40
0.9
0 30 60
fun
tion in equilibrium. At this levels the
urve of the
ase 3 Pareto
distribution fun
tion is above the
urve of the
ase 1 Pareto distribu-
tion fun
tion. We therefore observe the de
reasing P2 's. But when we
in
rease the initial
apital and thereby in
rease the survival probability
the pi
ture
hanges. At this level the fat tail is taking over, and we
get the desired ee
t namely that P2 in
reases when the tail is getting
fatter.
We now
on
lude that the simple OPC1 is easy to
al
ulate, and it
works reasonable. OPC2 is more
ompli
ated to
al
ulate and we have
to be more
areful when we use it, be
ause it is highly depended on
the initial
apital. But with a \suitable" level of the initial
apital we
have seen that it works well. It is not possible to say whether OPC1
or OPC2 is best in general.
6.3. Example 3. As seen above, it is not possible to rank the dierent
OPC's. But in this example we will try to do it for an agent that weights
all four OPC's equally and who has a known utility fun
tion.
Let the situation be as in example 1. We further assume that the
onsidered agent has an exponential utility fun
tion with risk aversion
oeÆ
ient = 0:03 and initial
apital u = 0.
HOW TO HEDGE UNKNOWN RISK 19
We will now evaluate how OPC1, OPC2, OPC3 and OPC4 are do-
ing in relation of having equal risk quantity (RQ), equal ruin proba-
bility (RP), equal expe
ted utility (EU) and maximal expe
ted utility
(MEU). This is done in the following way. Let RQij denote the risk
quantity in state j if Pj is determined from OPCi. We then introdu
e
the variable RQi to denote how well OPCi is doing in relation of having
equal risk quantities, let RQi be given by
P3
RQi = P4 jP =1 pj jRQ1j RQij j :
3
i=1 ( j =1 pj jRQ1j RQij j)
The numerator denotes how far the RQ is from the optimal RQ (the
RQ obtained by OPC1) weighted by the state probabilities. The de-
nominator is in
luded in order to normalize the expression su
h that
it
an be
ompared with the three other
ases. Note that RQi is 0 for
i = 1.
A
orresponding formula is used to
al
ulate the variables that de-
notes how well the OPCi's are doing in relation of having equal ruin
probabilities (the RPi 's) and equal expe
ted utilities (the EUi 's). For
the maximal expe
ted utility we use the following formula
MEUi = P4
jMEU4 MEUij
i=1 (jMEU4 MEUi j)
In the table below we have listed RQi , RPi , EUi and MEUi for
i=1,2,3 and 4.
OPC number RQ RP EU MEU
1 0 0.300 0.113 0.468 0.881
2 0.144 0 0.220 0.190 0.554
3 0.105 0.195 0 0.342 0.643
4 0.751 0.504 0.667 0 1.922
In the
olumn to the right the sum of the gures are listed. From
this
olumn we
an now rank the OPC's. It is seen that OPC2 is
the best OPC to use for the agent
onsidered in this example. But
it is important to note that the analysis is highly depended on the
preferen
es of the agent. If we set the risk aversion
oeÆ
ient dierent
20 C. VORM CHRISTENSEN
state premiums as possible in the least square sense, i.e. the least
square strategy is obtained by the following portfolio of se
urities.
2
P1 3
m = (C T C ) 1 C T 4 ... 5
Pn
The insuran
e
ompany would of
ourse prefer to follow the optimal
trading strategy given by Pi = E [Li ℄ + Æi but this is impossible in this
market. But if it has been possible the insuran
e
ompany would have
been willing to pay more for the optimal strategy than for the least
square strategy. Therefore if the insuran
e
ompany follows the least
square strategy they should
harge a premium that is larger than the
pri
e of the least square strategy. They thereby get a
ompensation for
not having the optimal strategy but only the least square strategy.
7.2. The restri
ted premium
ase. Let us now as in the
omplete
ase
onsider the situation where the insuran
e
ompany is unable to
harge the desired premium be
ause of some
ompetitive reasons. We
again set the possible xed premium that
an be
harged to P0 .
The problem now is that we want to set the Pi 's a
ording to OPC1,
OPC2, OPC3 or OPC4 but the equation P0 = P1 q1 + + Pn qn is no
longer valid. We are no longer able to
onstru
t the n AD se
urities
in this in
omplete market. But instead of
hoosing the Pi s a
ording
to OPC1, OPC2, OPC3 or OPC4 we
ould
hoose the
orresponding
least square solution. We then just have to repla
e the equation P0 =
P1 q1 + + Pn qn with an equation that makes sure that the pri
e of the
least square portfolio is equal to P0 . Before we solve this problem we
make the following denition in relation to OPC1, OPC2 and OPC3.
We return to OPC4 later.
Denition 7.2. The in
omplete optimal premium
hoi
e (IOPC)
is dened as a
hoi
e of premiums (Pi 's) su
h that
The Pi's solve the n 1 rst equations in OPC1, OPC2 or OPC3.
The pri
e of the least square strategy
orresponding to the Pi's is
P0 .
The IOPC is then found by the following theorem
Theorem 7.3. The solution to the IOPC
orresponding to OPC1,
OPC2 or OPC3 is found by solving the n equations with n unknowns
from the OPC. But where the last equation in OPC1, OPC2 and OPC3
is repla
ed by
2
P1 3
P0 = q~ 4 ... 5 = q~1 P1 + + q~n Pn
Pn
22 C. VORM CHRISTENSEN
where q~ is given by
q~ = v (C T C ) 1 C T
Proof. The least square strategy
orresponding to (P1 ; : : : ; Pn ) is given
by
2
P1 3
m = (C T C ) 1 C T 4 ... 5
Pn
and the pri
e of the state se
urities is given by v = (v1 ; : : : ; vk ). The
pri
e of the least square strategy
orresponding to (P1 ; : : : ; Pn) is then
given by
2
P1 3
vm = v (C T C ) 1 C T 4 ... 5
Pn
2 3
P1
= q~ ... 5 = q~1 P1 + + q~n Pn
4 (7.4)
Pn
where q~ is given by
q~ = v (C T C ) 1 C T
It then follows by denition 7.2 that if the last equation in OPC1,
OPC2 or OPC3 is repla
ed by 7.4, we will obtain the
orresponding
IOPC.
For OPC4 the situation is dierent. OPC4 is a maximization prob-
lem and we therefore just have to reformulate the problem to a problem
of dimension k instead of a problem of dimension n. The IOPC for
OPC4 therefore takes the following form
max
m
E [v (u +
i m Li )℄
st vm = P0 :
The maximization problem is solved in the same way as des
ribed in
se
tion 5.
8. Con
lusion
In this paper we have been looking at risk with more than one prior
estimate of the frequen
y. As mentioned in the introdu
tion it is not
possible to hedge this kind of risk using traditional insuran
e pra
ti
e
only, so a new method was
alled for. In this paper we present a
model that is able to manage this kind of risk. The model works by
using traditional insuran
e pra
ti
e and trading in the se
urity market
HOW TO HEDGE UNKNOWN RISK 23
simultaneously. The paper shows how this new method works both in
omplete and in
omplete markets.
Further we
onsider the
ase where the premium the insuran
e
om-
pany
an
harge is restri
ted. In this
ase the insuran
e
ompany has
to
hoose an allo
ation of the restri
ted premium
orresponding to the
states of the world. We propose four dierent methods of solving this
problem. These four methods are then analysed and evaluated and by
examples advantages and disadvantages are illustrated. We also show
a way to rank the four methods in the
ase where we
onsider an agent
that evaluates the four OPC equally and has a known utility fun
tion.
Referen
es
1. Beauregard, R.A. and Fraleigh, J.B. (1990), Linear Algebra, 2nd Edition,
Addison-Wesley Publishing Company.
2. Chi
hilnisky, G. and Heal, G. (1998), Managing Unknown Risks, The future of
global reinsuran
e, The Journal of Portfolio Management, pp 85-91, summer
1998.
3. Henderson, J.M. and Quandt, R.E. (1980) Mi
roe
onomi
Theory, A Mathe-
mati
al Approa
h, M
Graw-Hill book
ompany.
(C. Vorm Christensen) Department of Theoreti
al Statisti
s and Op-
erations Resear
h, University of Aarhus, Ny Munkegade 116, 8000
Aarhus C, Denmark
E-mail address : vormimf.au.dk