Beruflich Dokumente
Kultur Dokumente
37{50
Abstra
t
In this paper is dis
ussed a simulation method developed by S.
Asmussen
alled
onjugate pro
esses whi
h is based on a version
of Wald's fundamental identity. With this method it is possible
to simulate within nite time risk reserve pro
esses with innite
time horizons. This allows us to
onstru
t Monte Carlo estimators
for the ruin probability, whi
h is one of the main problems in
insuran
e risk theory. Some examples of the Poisson/Exponential
and Poisson/Uniform
ases are presented.
1 Introdu
tion
One of the main problems in insuran
e risk theory is to estimate the
ruin probability [1-8,11℄. It
an be roughly des
ribed as follows.
The risk reserve pro
ess over (0; t℄ is the dieren
e between a pre-
mium deterministi
pro
ess u +
t and the a
umulated
laims Zt (a
ompound Poisson pro
ess), for some given initial
apital u 0 . The
premium in
ome rate
is xed by the insuran
e
ompany and is in-
dependent of t. The idea is to study the behavior of the risk reserve
pro
ess that models the a
umulated
apital over nite or innite time
Resear
h partially supported by a CONACyT s
holarship. This paper is part of
UAM-Iztapalapa.
1
Professor at Departamento de Sistemas, UAM-Az
apotzal
o.
37
38 Luis F. Hoyos-Reyes
horizons, in parti
ular the probability that exists a moment when the
risk pro
ess is negative. This is
alled the ruin probability.
The main purpose of this paper is to introdu
e Monte Carlo esti-
mators (MCEs) for the ruin probability in innite time horizon using
onjugate pro
esses. This approa
h allows us to simulate within nite
time a risk reserve pro
ess with innite time horizon. Here, we
onstru
t
a MCE for the ruin probability using the empiri
al distribution of the
ruin events after a suÆ
iently large number of simulations. Some ex-
amples of the Poisson/Exponential (P/E) and Poisson/Uniform (P/U)
ases are presented.
This paper is organized as follows. We begin in x2 by introdu
ing
basi
terminology and notation. In x3 we show a formulation for the
onjugate pro
ess and
onstru
t the MCEs. In x4 we
ompute exam-
ples of the P/E and P/U
ases. Finally, x5 presents some
on
luding
remarks.
2 Preliminaries
Assumption 2.1
(a) The
laims arrive a
ording to a Poisson pro
ess fNtgt0 with
intensity and inter
laim times fTt gt1 .
(b) The
laim sizes X1; X2 ; : : : are i.i.d nonnegative random variables
with a nite mean .
(
) Xi and fNt gt0 are independent.
Denition 2.2 The a
umulated
laim pro
ess is Zt := Nn=0 Xn for
P t
t 0, with X0 := 0.
We next re
all the
lassi
al risk reserve pro
ess [2,7,8℄.
Denition 2.3 Let u be the initial
apital and
> 0 be the premium
in
ome rate.
We dene the risk reserve pro
ess
Yt := Zt
t; t 2 (0; 1);
and the time to ruin
:= inf ft > 0 : Yt > ug :
Insuran
e ruin problems 39
and if for some xed 0 2 the parameter set
ontains all 2 R for
whi
h (1) denes a probability density for some h (). 0
Z 1 x
( ) = e dF (x)
1
Z 1 x e( 0 )x
= e dF (x)
1 E0 e( 0 )X 0
= ((+ )0) :
0
0 0
EeYt = Zt ( )=e
t :
Hen
e, from (2) we have
EeYt = et(X () 1)
t :
Applying the log fun
tion to both sides of the latter equation and di-
viding by t,
ompletes the proof.
Now from Proposition 2.5
E eZt = E e(+ )Zt =E e( )Zt :
0
0
0
0
Using (2)
E eZt = et ( )( () 1) ;
0 0
whi
h implies that under P , Zt is also a
ompound Poisson pro
ess with
arrival rate = ( 0) and
laims distribution F . Therefore,
0
repla
ing E with E in Proposition 2.6 we obtain
0X ( 0 ) = = = 0 ( 0 ); 0
whi
h yields
0 ( 0 ) = 0:
0
and, therefore,
Using (4)
= 0 (
0
0 );
and
0 = 00 ( 0 ) = 0:
The last two equalities and the
onvexity of () yield the desired
0
on
lusion.
3 Monte Carlo estimators
Our main purpose in this se
tion is to estimate the ruin probability
onsidering an innite time horizon for the risk reserve pro
ess. We
rst introdu
e some denitions.
Denition 3.1 Let u and be as in Denition 2.3. The ruin probabil-
ity in nite time is
(u; T ) := P ( < T );
and the ruin probability in innite time is
(u) := P ( < 1):
E (Zt ) = tE X:
Insuran
e ruin problems 43
Now, from Proposition 2.7 we have E (Zt
t) > 0 and, therefore, < 0,
and using (5)
lim 1t (Zt
t!1
tE X tE X ) = E X > 0 a.s.
This implies
(Zt t(1 + )E X ) ! +1 a.s.;
whi
h
ompletes the proof.
Consider a
onjugate family (F )2 governing a random walk fSt gt0
in dis
rete or
ontinuous time. Dene FT := (St ; t T ), with the usual
extension to stopping times.
Next, we present the version of the Wald's fundamental identity used
by Asmussen [1,2℄. The proof
an be seen in [3℄.
Theorem 3.3 Let be a stopping time for fSt gt0 and G 2 F , G
f < 1g. Then for ea
h 0; 2
1X
n
Ri ! (u) as n ! 1;
n i=1
where Ri is the nal value of the realization of the
onjugate pro
ess
after simulation i (i = 1; 2; : : : ).
21 (1+) = E1 [exp f 2(
+ 1)Y + (1 )g I f < T g℄ 2(u; T );
1
(d) The overshot B (u) of the risk pro
ess, dened as B (u) := Y u is
useful to
al
ulate 2 . It is known [1,2℄ that when the
laims are
exponentially distributed and the arrival pro
ess is Poisson (P/E
1
The Poisson/Exponential
ase has been extensively resear
hed [1-5℄ be-
ause it is easy to
al
ulate the ruin probability for the innite time
46 Luis F. Hoyos-Reyes
Note that the dieren
e between the varian
es is signi
ant, whi
h is
an statisti
al advantage [9℄ to
onstru
t
onden
e intervals for (u).
To show some numeri
al results, let = 1, = 0:8, = 0:1,
= 0:88.
>From Proposition 2.6, we
an see that Nt is a Poisson pro
ess with
arrival rate
, and X is exponentially distributed with parameter =
.
Insuran
e ruin problems 47
Let us assume that the
laims size distribution is uniform over (0; 1).
First of all, we need to nd the distibution F , and then we have to
show an expression for the
onjugate pro
ess RT (1 ) . From (1)
1
e( )x 1
F (x) =
1 ; 0 < x < 1:
0
e 0
Re
all that under P , Zt is also a
ompound Poisson pro
ess with pa-
rameter
= e
+ 1 =(
+ 1 ):
1
From Proposition 2.6 and (8) we obtain that in the nite horizon
ase,
and letting
0 :=
+ 1,
! !
e
+1 1
(11) RT1 (1+) = exp
0 Y +
0
1
0
I f < T g;
whereas in the innite time horizon
48 Luis F. Hoyos-Reyes
! !
e
+1 1
(12) R1 (1+) = exp
0 Y +
0
1
0
:
Departamento de Sistemas ,
UAM - Az apotzal o,
02200 M
exi
o D.F., MEXICO,
Referen
es
[1℄ Asmussen, S., Approximations for the probability of ruin within
nite time, S
andinavian A
tuarial J. 20 (1984), 31-57.
[2℄ Asmussen, S., Conjugate pro
esses and the simulation of ruin
problems, Sto
hasti
Pro
esses and their Appli
ations 20 (1985),
213-229.
[3℄ Asmussen, S., Applied Probability and Queues, Wiley, Chi
hester,
U.K., 1987.
[4℄ Asmussen, S. and Rolski, T., Computational methods in risk the-
ory: A matrix-algorithmi
approa
h, Insuran
e: Mathemati
s and
E
onomi
s 10 (1991), 259-274.
[5℄ Beard, R.E., Pentikainen, T. and Pessonen, E., Risk Theory,
Chapman and Hall, New York, 1984.
[6℄ Dassios, A. and Embre
hts, P., Martingales and insuran
e risk,
Commun. Statist.-Sto
hasti
Models 5 (1989), 181-217.
[7℄ Embre
hts, P., Sto
hasti
Modelling in insuran
e, CLAPEM-IV
Pro
eedings, Mexi
o City 1990.
[8℄ Embre
hts, P. and Wouters, P., Simulating risk solven
y, Insur-
an
e:Mathemati
s and E
onomi
s 9 (1990), 141-148.
[9℄ Ross, S.M., Sto
hasti
Pro
esses, Wiley, New York, 1983.
[10℄ Ross, S.M., A Course in Simulation, Ma
millan, New York, 1990.
50 Luis F. Hoyos-Reyes