Professional Documents
Culture Documents
Risk Theory
Risk Theory
Risk Theory
by Hanspeter Schmidli
Institute of Mathematics
University of Cologne
CONTENTS i
Contents
1. Risk Models 1
1.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2. The Compound Binomial Model . . . . . . . . . . . . . . . . . . . . . 2
1.3. The Compound Poisson Model . . . . . . . . . . . . . . . . . . . . . . 3
1.4. The Compound Mixed Poisson Model . . . . . . . . . . . . . . . . . . 6
1.5. The Compound Negative Binomial Model . . . . . . . . . . . . . . . 7
1.6. A Note on the Individual Model . . . . . . . . . . . . . . . . . . . . . 7
1.7. A Note on Reinsurance . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.7.1. Proportional reinsurance . . . . . . . . . . . . . . . . . . . . . 9
1.7.2. Excess of loss reinsurance . . . . . . . . . . . . . . . . . . . . 9
1.8. Computation of the Distribution of S in the Discrete Case . . . . . . 11
1.9. Approximations to S . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.9.1. The Normal Approximation . . . . . . . . . . . . . . . . . . . 14
1.9.2. The Translated Gamma Approximation . . . . . . . . . . . . 15
1.9.3. The Edgeworth Approximation . . . . . . . . . . . . . . . . . 17
1.9.4. The Normal Power Approximation . . . . . . . . . . . . . . . 19
1.10. Premium Calculation Principles . . . . . . . . . . . . . . . . . . . . . 21
1.10.1. The Expected Value Principle . . . . . . . . . . . . . . . . . . 21
1.10.2. The Variance Principle . . . . . . . . . . . . . . . . . . . . . . 22
1.10.3. The Standard Deviation Principle . . . . . . . . . . . . . . . . 22
1.10.4. The Modied Variance Principle . . . . . . . . . . . . . . . . 22
1.10.5. The Principle of Zero Utility . . . . . . . . . . . . . . . . . . 22
1.10.6. The Mean Value Principle . . . . . . . . . . . . . . . . . . . . 24
1.10.7. The Exponential Principle . . . . . . . . . . . . . . . . . . . . 25
1.10.8. The Esscher Principle . . . . . . . . . . . . . . . . . . . . . . 27
1.10.9. The Proportional Hazard Transform Principle . . . . . . . . . 27
1.10.10.The Percentage Principle . . . . . . . . . . . . . . . . . . . . 27
1.10.11.Desirable Properties . . . . . . . . . . . . . . . . . . . . . . . 28
ii CONTENTS
2. Utility Theory 29
2.1. The Expected Utility Hypothesis . . . . . . . . . . . . . . . . . . . . 29
2.2. The Zero Utility Premium . . . . . . . . . . . . . . . . . . . . . . . . 32
2.3. Optimal Insurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.4. The Position of the Insurer . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5. Pareto-Optimal Risk Exchanges . . . . . . . . . . . . . . . . . . . . . 36
3. Credibility Theory 41
3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2. Bayesian Credibility . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2.1. The Poisson-Gamma Model . . . . . . . . . . . . . . . . . . . 43
3.2.2. The Normal-Normal Model . . . . . . . . . . . . . . . . . . . 44
3.2.3. Is the Credibility Premium Formula Always Linear? . . . . . 45
3.3. Empirical Bayes Credibility . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.1. The B uhlmann Model . . . . . . . . . . . . . . . . . . . . . . 46
3.3.2. The B uhlmann Straub model . . . . . . . . . . . . . . . . . . 50
3.3.3. The B uhlmann Straub model with missing data . . . . . . . . 56
3.4. General Bayes Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.5. Hilbert Space Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4. The Cramer-Lundberg Model 62
4.1. Denition of the Cramer-Lundberg Process . . . . . . . . . . . . . . . 62
4.2. A Note on the Model and Reality . . . . . . . . . . . . . . . . . . . . 63
4.3. A Dierential Equation for the Ruin Probability . . . . . . . . . . . . 64
4.4. The Adjustment Coecient . . . . . . . . . . . . . . . . . . . . . . . 67
4.5. Lundbergs Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.6. The Cramer-Lundberg Approximation . . . . . . . . . . . . . . . . . 72
4.7. Reinsurance and Ruin . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.7.1. Proportional Reinsurance . . . . . . . . . . . . . . . . . . . . 74
4.7.2. Excess of Loss Reinsurance . . . . . . . . . . . . . . . . . . . 75
4.8. The Severity of Ruin and the Distribution of infC
t
: t 0 . . . . . 77
CONTENTS iii
4.9. The Laplace Transform of . . . . . . . . . . . . . . . . . . . . . . . 79
4.10. Approximations to . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.10.1. Diusion Approximations . . . . . . . . . . . . . . . . . . . . 83
4.10.2. The deVylder Approximation . . . . . . . . . . . . . . . . . . 85
4.10.3. The Beekman-Bowers Approximation . . . . . . . . . . . . . . 87
4.11. Subexponential Claim Size Distributions . . . . . . . . . . . . . . . . 88
4.12. The Time of Ruin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.13. Seals Formulae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.14. Finite Time Lundberg Inequalities . . . . . . . . . . . . . . . . . . . . 97
5. The Renewal Risk Model 101
5.1. Denition of the Renewal Risk Model . . . . . . . . . . . . . . . . . . 101
5.2. The Adjustment Coecient . . . . . . . . . . . . . . . . . . . . . . . 102
5.3. Lundbergs Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.3.1. The Ordinary Case . . . . . . . . . . . . . . . . . . . . . . . . 106
5.3.2. The General Case . . . . . . . . . . . . . . . . . . . . . . . . 109
5.4. The Cramer-Lundberg Approximation . . . . . . . . . . . . . . . . . 109
5.4.1. The Ordinary Case . . . . . . . . . . . . . . . . . . . . . . . . 109
5.4.2. The general case . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.5. Diusion Approximations . . . . . . . . . . . . . . . . . . . . . . . . 112
5.6. Subexponential Claim Size Distributions . . . . . . . . . . . . . . . . 113
5.7. Finite Time Lundberg Inequalities . . . . . . . . . . . . . . . . . . . . 116
6. Some Aspects of Heavy Tails 118
6.1. Why Heavy Tailed Distributions Are Dangerous . . . . . . . . . . . . 118
6.2. How to Find Out Whether a Distribution Is Heavy Tailed or Not . . 121
6.2.1. The Mean Residual Life . . . . . . . . . . . . . . . . . . . . . 121
6.2.2. How to Estimate the Mean Residual Life Function . . . . . . 122
6.3. Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.3.1. The Exponential Distribution . . . . . . . . . . . . . . . . . . 125
6.3.2. The Gamma Distribution . . . . . . . . . . . . . . . . . . . . 125
iv CONTENTS
6.3.3. The Weibull Distribution . . . . . . . . . . . . . . . . . . . . 126
6.3.4. The Lognormal Distribution . . . . . . . . . . . . . . . . . . . 127
6.3.5. The Pareto Distribution . . . . . . . . . . . . . . . . . . . . . 127
6.3.6. Parameter Estimation for the Fire Insurance Data . . . . . . 129
6.4. Verication of the Chosen Model . . . . . . . . . . . . . . . . . . . . 131
6.4.1. Q-Q-Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.4.2. The Fire Insurance Example . . . . . . . . . . . . . . . . . . . 135
7. The Ammeter Risk Model 138
7.1. Mixed Poisson Risk Processes . . . . . . . . . . . . . . . . . . . . . . 138
7.2. Denition of the Ammeter Risk Model . . . . . . . . . . . . . . . . . 139
7.3. Lundbergs Inequality and the Cramer-Lundberg Approximation . . . 141
7.3.1. The Ordinary Case . . . . . . . . . . . . . . . . . . . . . . . . 141
7.3.2. The General Case . . . . . . . . . . . . . . . . . . . . . . . . 146
7.4. The Subexponential Case . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.5. Finite Time Lundberg Inequalities . . . . . . . . . . . . . . . . . . . . 154
8. Change of Measure Techniques 156
8.1. The Cramer-Lundberg Case . . . . . . . . . . . . . . . . . . . . . . . 157
8.2. The Renewal Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
8.2.1. Markovization Via the Time Since the Last Claim . . . . . . . 161
8.2.2. Markovization Via the Time Till the Next Claim . . . . . . . 163
8.3. The Ammeter Risk Model . . . . . . . . . . . . . . . . . . . . . . . . 164
9. The Markov Modulated Risk Model 168
9.1. Denition of the Markov Modulated Risk Model . . . . . . . . . . . . 168
9.2. The Lundberg Exponent and Lundbergs Inequality . . . . . . . . . . 170
9.3. The Cramer-Lundberg Approximation . . . . . . . . . . . . . . . . . 175
9.4. Subexponential Claim Sizes . . . . . . . . . . . . . . . . . . . . . . . 177
9.5. Finite Time Lundberg Inequalities . . . . . . . . . . . . . . . . . . . . 178
A. Stochastic Processes 180
A.1. Bibliographical Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 181
CONTENTS v
B. Martingales 182
B.1. Bibliographical Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 182
C. Renewal Processes 183
C.1. Poisson Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
C.2. Renewal Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
C.3. Bibliographical Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 191
D. Brownian Motion 192
E. Random Walks and the Wiener-Hopf Factorization 193
E.1. Bibliographical Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 195
F. Subexponential Distributions 196
F.1. Bibliographical Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 201
G. Concave and Convex Functions 202
Table of Distribution Functions 206
References 208
Index 214
vi CONTENTS
1. RISK MODELS 1
1. Risk Models
1.1. Introduction
Let us consider a (collective) insurance contract in some xed time period (0, T], for
instance T = 1 year. Let N denote the number of claims in (0, T] and Y
1
, Y
2
, . . . , Y
N
the corresponding claims. Then
S =
N
i=1
Y
i
is the accumulated sum of claims. We assume
i) The claim sizes are not aected by the number of claims.
ii) The amount of an individual claim is not aected by the others.
iii) All individual claims are exchangeable.
In probabilistic terms:
i) N and (Y
1
, Y
2
, . . .) are independent.
ii) Y
1
, Y
2
, . . . are independent.
iii) Y
1
, Y
2
, . . . have the same distribution function, G say.
We assume that G(0) = 0, i.e. the claim amounts are positive. Let M
Y
(r) = II E[e
rY
i
],
n
= II E[Y
n
1
] if the expressions exist and =
1
. The distribution of S can be written
as
II P[S x] = II E[II P[S x [ N]] =
n=0
II P[S x [ N = n]II P[N = n]
=
n=0
II P[N = n]G
n
(x) .
This is in general not easy to compute. It is often enough to know some character-
istics of a distribution.
II E[S] = II E
_
N
i=1
Y
i
_
= II E
_
II E
_
N
i=1
Y
i
N
__
= II E
_
N
i=1
_
= II E[N] = II E[N]
2 1. RISK MODELS
and
II E[S
2
] = II E
_
II E
__
N
i=1
Y
i
_
2
N
__
= II E
_
II E
_
N
i=1
N
j=1
Y
i
Y
j
N
__
= II E[N
2
+N(N 1)
2
] = II E[N
2
]
2
+ II E[N](
2
2
) ,
and thus
Var[S] = Var[N]
2
+ II E[N] Var[Y
1
] .
The moment generating function of S becomes
M
S
(r) = II E[e
rS
] = II E
_
exp
_
r
N
i=1
Y
i
__
= II E
_
N
i=1
e
rY
i
_
= II E
_
II E
_
N
i=1
e
rY
i
N
__
= II E
_
N
i=1
M
Y
(r)
_
= II E
_
(M
Y
(r))
N
= II E
_
e
N log(M
Y
(r))
= M
N
(log(M
Y
(r)))
where M
N
(r) is the moment generating function of N. The coecient of skewness
II E[(S II E[S])
3
]/(Var[S])
3/2
can be computed from the moment generating function
by using the formula
II E[(S II E[S])
3
] =
d
3
dr
3
log(M
S
(r))
r=0
.
1.2. The Compound Binomial Model
We make in addition the following assumptions:
The whole risk consists of several independent exchangeable single risks.
The time interval under consideration can be split into several independent and
exchangeable smaller intervals.
There is at most one claim per single risk and interval.
Let the probability of a claim in a single risk in an interval be p. Then it follows
from the above assumptions that
N B(n, p)
for some n IIN. We obtain
II E[S] = np,
1. RISK MODELS 3
Var[S] = np(1 p)
2
+np(
2
2
) = np(
2
p
2
)
and
M
S
(r) = (pM
Y
(r) + 1 p)
n
.
Let us consider another characteristics of the distribution of S, the skewness. We
need to compute II E[(S II E[S])
3
].
d
3
dr
3
nlog(pM
Y
(r) + 1 p) = n
d
2
dr
2
_
pM
t
Y
(r)
pM
Y
(r) + 1 p
_
= n
d
dr
_
pM
tt
Y
(r)
pM
Y
(r) + 1 p
p
2
M
t
Y
(r)
2
(pM
Y
(r) + 1 p)
2
_
= n
_
pM
ttt
Y
(r)
pM
Y
(r) + 1 p
3p
2
M
tt
Y
(r)M
t
Y
(r)
(pM
Y
(r) + 1 p)
2
+
2p
3
(M
t
Y
(r))
3
(pM
Y
(r) + 1 p)
3
_
.
For r = 0 we get
II E[(S II E[S])
3
] = n(p
3
3p
2
2
+ 2p
3
3
)
from which the coecient of skewness can be computed.
Example 1.1. Assume that the claim amounts are deterministic, y
0
say. Then
II E[(S II E[S])
3
] = ny
3
0
(p 3p
2
+ 2p
3
) = 2ny
3
0
p(
1
2
p)(1 p) .
Thus
II E[(S II E[S])
3
] 0 p
1
2
.
2
) =
2
and
M
S
(r) = exp(M
Y
(r) 1) .
Let us also compute the coecient of skewness.
d
3
dr
3
log(M
S
(r)) =
d
3
dr
3
((M
Y
(r) 1)) = M
ttt
Y
(r)
and thus
II E[(S II E[S])
3
] =
3
.
The coecient of skewness is
II E[(S II E[S])
3
]
(Var[S])
3/2
=
3
_
3
2
> 0 .
Problem: The distribution of S is always positively skewed.
Example 1.2. (Fire insurance) An insurance company models claims caused
by re as LN(m,
2
). Let us rst compute the moments
n
= II E[Y
n
1
] = II E[e
nlog Y
1
] = M
log Y
1
(n) = exp
2
n
2
/2 +nm .
Thus
II E[S] = exp
2
/2 +m ,
Var[S] = exp2
2
+ 2m
and
II E[(S II E[S])
3
]
(Var[S])
3/2
=
exp9
2
/2 + 3m
_
exp6
2
+ 6m
=
exp3
2
/2
The computation of the characteristics of a risk is much easier for the compound
Poisson model than for the compound binomial model. But using a compound
Poisson model has another big advantage. Assume that a portfolio consists of several
independent single risks S
(1)
, S
(2)
, . . . , S
(j)
, each modelled as compound Poisson. For
1. RISK MODELS 5
simplicity we use j = 2 in the following calculation. We want to nd the moment
generating function of S
(1)
+S
(2)
.
M
S
(1)
+S
(2) (r) = M
S
(1) (r)M
S
(2) (r)
= exp
(1)
(M
Y
(1) (r) 1) exp
(2)
(M
Y
(2) (r) 1)
= exp
_
(1)
M
Y
(1) (r) +
(2)
M
Y
(2) (r) 1
__
where =
(1)
+
(2)
. It follows that S
(1)
+ S
(2)
is compound Poisson distributed
with Poisson parameter and claim size distribution function
G(x) =
(1)
G
(1)
(x) +
(2)
G
(2)
(x) .
The claim size can be obtained by choosing it from the rst risk with probability
(1)
/ and from the second risk with probability
(2)
/.
Let us now split the claim amounts into dierent classes. Let A
1
, A
2
, . . . , A
m
be some disjoint sets with II P[Y
1
m
k=1
A
k
] = 1. Let p
k
= II P[Y
1
A
k
] be the
probability that a claim is in claim size class k. We can assume that p
k
> 0 for all
k. We denote by N
k
the number of claims in claim size class k. Because the claim
amounts are independent it follows that, given N = n, the vector (N
1
, N
2
, . . . , N
m
)
is conditionally multinomial distributed with parameters p
1
, p
2
, . . . , p
m
, n. We now
want to nd the unconditioned distribution of (N
1
, N
2
, . . . , N
m
). Let n
1
, n
2
, . . . , n
m
be natural numbers and n = n
1
+n
2
+ +n
m
.
II P[N
1
= n
1
, N
2
= n
2
, . . . , N
m
= n
m
] = II P[N
1
= n
1
, N
2
= n
2
, . . . , N
m
= n
m
, N = n]
= II P[N
1
= n
1
, N
2
= n
2
, . . . , N
m
= n
m
[ N = n]II P[N = n]
=
n!
n
1
!n
2
! n
m
!
p
n
1
1
p
n
2
2
p
nm
m
n
n!
e
=
m
k=1
(p
k
)
n
k
n
k
!
e
p
k
It follows that N
1
, N
2
, . . . , N
m
are independent and that N
k
is Pois(p
k
) distributed.
Because the claim sizes are independent of N the risks
S
k
=
N
i=1
Y
i
1I
Y
i
A
k
n
n!
e
_
=
_
0
n
n!
e
dH() .
The moments are
II E[S] = II E[II E[S [ ]] = II E[] = II E[], (1.1a)
II E[S
2
] = II E[II E[S
2
[ ]] = II E[
2
+
2
2
] = II E[
2
]
2
+ II E[]
2
(1.1b)
and
II E[S
3
] = II E[
3
+ 3
2
2
+
3
3
] = II E[
3
]
3
+ 3II E[
2
]
2
+ II E[]
3
. (1.1c)
Thus the variance is
Var[S] = Var[]
2
+ II E[]
2
and the third centralized moment becomes
II E[(S II E[])
3
]
= II E[
3
]
3
+ 3II E[
2
]
2
+ II E[]
3
3(II E[
2
]
2
+ II E[]
2
)II E[] + 2II E[]
3
3
= II E[( II E[])
3
]
3
+ 3 Var[]
2
+ II E[]
3
.
We can see that the coecient of skewness can also be negative.
It remains to compute the moment generating function
M
S
(r) = II E
_
II E
_
e
rS
[
= II E[exp(M
Y
(r) 1)] = M
(M
Y
(r) 1) . (1.2)
1. RISK MODELS 7
1.5. The Compound Negative Binomial Model
Let us rst consider an example of the compound mixed Poisson model. Assume
that (, ). Then
M
N
(r) = M
(e
r
1) =
_
(e
r
1)
_
=
_
+1
1 (1
+1
)e
r
_
.
In this case N has a negative binomial distribution. Actuaries started using the
negative binomial distribution for the number of claims a long time ago. They
recognized that a Poisson distribution seldomly ts the real data. Estimating the
parameters in a negative binomial distribution yielded satisfactory results.
Let now N NB(, p). Then
II E[S] =
(1 p)
p
and
Var[S] =
(1 p)
p
2
2
+
(1 p)
p
(
2
2
) .
The moment generating function is
M
S
(r) =
_
p
1 (1 p)M
Y
(r)
_
.
The third centralized moment can be computed as
II E[(S II E[S])
3
] =
_
1
p
1
_
3
+ 3
_
1
p
1
_
2
2
+ 2
_
1
p
1
_
3
3
.
Note that the compound negative binomial distribution is always positively skewed.
Thus the compound negative binomial distribution does not full all desired prop-
erties. Nevertheless in practice almost all risks are positively skewed.
1.6. A Note on the Individual Model
Assume that a portfolio consist of m independent individual contracts (S
(i)
)
im
.
There can be at most one claim for each of the contracts. Such a claim occurs with
probability p
(i)
. Its size has distribution function F
(i)
and moment generating func-
tion M
(i)
(r). Let =
m
i=1
p
(i)
. The moment generating function of the aggregate
claims from the portfolio is
M
S
(r) =
m
i=1
(1 +p
(i)
(M
(i)
(r) 1)) .
8 1. RISK MODELS
The term p
(i)
(M
(i)
(r) 1) is small for r not too large. Consider the logarithm
log(M
S
(r)) =
m
i=1
log(1 +p
(i)
(M
(i)
(r) 1))
i=1
p
(i)
(M
(i)
(r) 1) =
_
m
i=1
p
(i)
(M
(i)
(r) 1)
_
.
The last expression turns out to be the logarithm of the moment generating function
of a compound Poisson distribution. In fact this derivation is the reason that the
compound Poisson model is very popular amongst actuaries.
1.7. A Note on Reinsurance
For most of the insurance companies the premium volume is not big enough to carry
the risk. This is the case especially in the case of large claim sizes, like insurance
against damages caused by hurricanes or earthquakes. Therefore the insurers try to
share part of the risk with other companies. Sharing the risk is done via reinsurance.
Let S
I
denote the part of the risk taken by the insurer and S
R
the part taken by the
reinsurer. Reinsurance can act on the individual claims or it can act on the whole
risk S. Let f be an increasing function with f(0) = 0 and f(x) x for all x 0. A
reinsurance form acting on the individual claims is
S
I
=
N
i=1
f(Y
i
) , S
R
= S S
I
.
The most common reinsurance forms are
proportional reinsurance f(x) = x, (0 < < 1),
excess of loss reinsurance f(x) = minx, M, (M > 0).
We will consider these two reinsurance forms in the sequel. A reinsurance form
acting on the whole risk is
S
I
= f(S) , S
R
= S S
I
.
The most common example of this reinsurance form is the
stop loss reinsurance f(x) = minx, M, (M > 0).
1. RISK MODELS 9
1.7.1. Proportional reinsurance
For proportional reinsurance we have S
I
= S, and thus
II E[S
I
] = II E[S] ,
Var[S
I
] =
2
Var[S] ,
II E[(S
I
II E[S
I
])
3
]
(Var[S
I
])
3/2
=
II E[(S II E[S])
3
]
(Var[S])
3/2
and
M
S
I (r) = II E
_
e
rS
= M
S
(r) .
We can see that the coecient of skewness does not change. But the variance is
much smaller. The following consideration also suggest that the risk has decreased.
Let the premium charged for this contract be p and assume that the insurer gets p,
the reinsurer (1 )p. Let the initial capital of the company be u. The probability
that the company gets ruined after one year is
II P[S > p +u] = II P[S > p +u/] .
The eect for the insurer is thus like having a larger initial capital.
Remark. The computations for the reinsurer can be obtained by replacing
with (1 ).
1.7.2. Excess of loss reinsurance
Under excess of loss reinsurance we cannot give formulae of the above type for the
cumulants and the moment generating function. They all have to be computed from
the new distribution function of the claim sizes Y
I
i
= minY
i
, M. An indication
that the risk has decreased for the insurer is that the claim sizes are bounded. This
is wanted especially in the case of large claims.
Example 1.3. Let S be the compound Poisson model with parameter and
Pa(, ) distributed claim sizes. Assume that > 1, i.e. that II E[Y
i
] < . Let
10 1. RISK MODELS
us compute the expected value of the outgo paid by the insurer.
II E[Y
I
i
] =
_
M
0
x
( +x)
+1
dx +
_
M
M
( +x)
+1
dx
=
_
M
0
(x +)
( +x)
+1
dx
_
M
0
( +x)
+1
dx +M
( +M)
1
_
1
1
1
( +M)
1
_
+1
_
1
1
( +M)
_
+M
( +M)
=
1
( 1)( +M)
1
=
_
1
_
+M
_
1
_
1
=
_
1
_
+M
_
1
_
II E[Y
i
] .
For II E[S
I
] follows that
II E[S
I
] =
_
1
_
+M
_
1
_
II E[S] .
Let
N
R
=
N
i=1
1I
Y
i
>M
denote the number of claims the reinsurer has to pay for. We denote by q =
II P[Y
i
> M] the probability that a claim amount exceeds the level M. What is the
distribution of N
R
? We rst note that the moment generating function of 1I
Y
i
>M
is qe
r
+ 1 q.
i) Let N B(n, p). The moment generating function of N
R
is
M
N
R(r) = (p(qe
r
+ 1 q) + 1 p)
n
= (pqe
r
+ 1 pq)
n
.
Thus N
R
B(n, pq).
ii) Let N Pois(). The moment generating function of N
R
is
M
N
R(r) = exp((qe
r
+ 1 q) 1) = expq(e
r
1) .
Thus N
R
Pois(q).
iii) Let N NB(, p). The moment generating function of N
R
is
M
N
R(r) =
_
p
1 (1 p)(qe
r
+ 1 q)
_
=
_
p
p+qpq
1
_
1
p
p+qpq
_
e
r
_
.
Thus N
R
NB(,
p
p+qpq
).
1. RISK MODELS 11
1.8. Computation of the Distribution of S in the Discrete Case
Let us consider the case where the claim sizes (Y
i
) have an arithmetic distribution.
We can assume that II P[Y
i
IIN] = 1 by choosing the monetary unit appropriate. We
denote by p
k
= II P[N = k], f
k
= II P[Y
i
= k] and g
k
= II P[S = k]. For simplicity let us
assume that f
0
= 0. Let f
n
k
= II P[Y
1
+ Y
2
+ Y
n
= k] denote the convolutions of
the claim size distribution. Note that
f
(n+1)
k
=
k1
i=1
f
n
i
f
ki
.
We get the following identities:
g
0
= II P[S = 0] = II P[N = 0] = p
0
,
g
n
= II P[S = n] = II E[II P[S = n [ N]] =
n
k=1
p
k
f
k
n
.
We have explicit formulae for the distribution of S. But the computation of the
f
k
n
s is messy. An easier procedure is called for. Let us now make an assumption
on the distribution of N.
Assumption 1.4. Assume that there exist real numbers a and b such that
p
r
=
_
a +
b
r
_
p
r1
for r IIN 0.
Let us check the assumption for the distribution of the models we had considered.
i) Binomial B(n,p)
p
r
p
r1
=
n!
r!(nr)!
p
r
(1 p)
nr
n!
(r1)!(nr+1)!
p
r1
(1 p)
nr+1
=
(n r + 1)p
r(1 p)
=
p
1 p
+
(n + 1)p
r(1 p)
.
Thus
a =
p
1 p
, b =
(n + 1)p
1 p
.
12 1. RISK MODELS
ii) Poisson Pois()
p
r
p
r1
=
r
r!
e
r1
(r1)!
e
=
r
.
Thus
a = 0 , b = .
iii) Negative binomial NB(, p)
p
r
p
r1
=
(+r)
r!()
p
(1 p)
r
(+r1)
(r1)!()
p
(1 p)
r1
=
( +r 1)(1 p)
r
= 1 p +
( 1)(1 p)
r
.
Thus
a = 1 p , b = ( 1)(1 p) .
These are in fact the only distributions satisfying the assumption. If we choose
a = 0 we get by induction p
r
= p
0
b
r
/r!, which is the Poisson distribution.
If a < 0, then because a + b/r is negative for r large enough there must be an
n
0
IIN such that b = an
0
in order that p
r
0 for all r. Letting n = n
0
1 and
p = a(1 a) we get the binomial distribution.
If a > 0 we need a + b 0 in order that p
1
0. The case a + b = 0 can be
considered as the degenerate case of a Poisson distribution with = 0. Suppose
therefore that a +b > 0. In particular, p
r
> 0 for all r. Let k IIN 0. Then
k
r=1
rp
r
=
k
r=1
(ar +b)p
r1
= a
k
r=1
(r 1)p
r1
+ (a +b)
k
r=1
p
r1
= a
k1
r=1
rp
r
+ (a +b)
k1
r=0
p
r
.
This can be written as
kp
k
= (a 1)
k1
r=1
rp
r
+ (a +b)
k1
r=0
p
r
.
Suppose a 1. Then kp
k
(a +b)p
0
and therefore p
k
(a +b)p
0
/k. In particular,
1 p
0
=
r=1
p
r
(a +b)p
0
r=1
1
r
.
This is a contradiction. Thus a < 1. If we now let p = 1 a and = a
1
(a +b) we
get the negative binomial distribution.
We will use the following lemma.
1. RISK MODELS 13
Lemma 1.5. Let n 2. Then
i) II E
_
Y
1
i=1
Y
i
= r
_
=
r
n
,
ii) p
n
f
n
r
=
r1
k=1
_
a +
bk
r
_
f
k
p
n1
f
(n1)
rk
.
Proof. i) Noting that the (Y
i
) are iid. we get
II E
_
Y
1
i=1
Y
i
= r
_
=
1
n
n
j=1
II E
_
Y
j
i=1
Y
i
= r
_
=
1
n
II E
_
n
j=1
Y
j
i=1
Y
i
= r
_
=
r
n
.
ii) Note that f
(n1)
0
= 0. Thus
p
n1
r1
k=1
_
a +
bk
r
_
f
k
f
(n1)
rk
= p
n1
r
k=1
_
a +
bk
r
_
f
k
f
(n1)
rk
= p
n1
r
k=1
_
a +
bk
r
_
II P
_
Y
1
= k,
n
j=2
Y
j
= r k
_
= p
n1
r
k=1
_
a +
bk
r
_
II P
_
Y
1
= k,
n
j=1
Y
j
= r
_
= p
n1
r
k=1
_
a +
bk
r
_
II P
_
Y
1
= k
j=1
Y
j
= r
_
f
n
r
= p
n1
II E
_
a +
bY
1
r
j=1
Y
j
= r
_
f
n
r
= p
n1
_
a +
b
n
_
f
n
r
= p
n
f
n
r
.
n=1
p
n
f
n
r
= p
1
f
r
+
n=2
p
n
f
n
r
= (a +b)p
0
f
r
+
n=2
r1
i=1
_
a +
bi
r
_
f
i
p
n1
f
(n1)
ri
= (a +b)g
0
f
r
+
r1
i=1
_
a +
bi
r
_
f
i
n=2
p
n1
f
(n1)
ri
= (a +b)f
r
g
0
+
r1
i=1
_
a +
bi
r
_
f
i
g
ri
=
r
i=1
_
a +
bi
r
_
f
i
g
ri
.
14 1. RISK MODELS
These formulae are called Panjer recursion. Note that the convolutions have van-
ished. Thus a simple recursion can be used to compute g
r
.
Consider now the case where the claim size distribution is not arithmetic. Choose
a bandwidth > 0. Let
Y
i
= infn IIN : Y
i
n ,
S =
N
i=1
Y
i
and
Y
i
= supn IIN : Y
i
n , S =
N
i=1
Y
i
.
It is clear that
S S
S .
The Panjer recursion can be used to compute the distribution of
S and S. Note
that for S there is a positive probability of a claim of size 0. Thus the original
distribution of N has to be changed slightly. The corresponding new claim number
distribution can be found in Section 1.7.2 for the most common cases.
1.9. Approximations to S
We have seen that the distribution of S is rather hard to determine, especially if no
computers are available. Therefore approximations are called for.
1.9.1. The Normal Approximation
It is quite often the case that the number of claims is large. For a deterministic
number of claims we would use a normal approximation in such a case. So we
should try a normal approximation to S.
We t a normal distribution Z such that the rst two moments coincide. Thus
II P[S x] II P[Z x] =
_
x II E[N]
_
Var[N]
2
+ II E[N](
2
2
)
_
.
In particular for a compound Poisson distribution the formula is
II P[S x]
_
x
2
_
.
Example 1.6. Let S have a compound Poisson distribution with parameter = 20
and Pa(4,3) distributed claim sizes. Note that = 1 and
2
= 3. Figure 1.1 shows
1. RISK MODELS 15
10 20 30 40 50
0.01
0.02
0.03
0.04
0.05
Figure 1.1: Densities of S and its normal approximation
the densities of S (solid line) and of its normal approximation (dashed line).
Assume that we want to nd a premium p such that II P[S > p] 0.05. Using
the normal approximation we nd that
II P[S > p] 1
_
p 20
60
_
= 0.05 .
Thus
p 20
60
= 1.6449
or p = 32.7413. If the accumulated claims only should exceed the premium in 1%
of the cases we get
p 20
60
= 2.3263
or p = 38.0194. The exact values are 33.94 (5% quantile) and 42.99 (1% quantile).
The approximation for the 5% quantile is quite good, but the approximation for the
1% quantile is not satisfactory.
Problem: It often turns out that loss distributions are skewed. The normal
distribution is not skewed. Hence we cannot hope that the approximation ts well.
We need an additional parameter to t to the distribution.
1.9.2. The Translated Gamma Approximation
The idea of the translated gamma approximation is to approximate S by k + Z
where k is a real constant and Z has a gamma distribution (, ). The rst three
16 1. RISK MODELS
moments should coincide. Note that II E[((k +Z) II E[k +Z])
3
] = 2
3
. Denote by
m = II E[S] the expected value of the risk, by
2
its variance and by its coecient
of skewness. Then
k +
= m,
2
=
2
and
2
3
3/2
=
2
= .
Solving this system yields
=
4
2
, =
=
2
, k = m
= m
2
.
Remarks.
i) Note that > 0 must be fullled.
ii) It may happen that k becomes negative. In that case the approximation has (as
the normal approximation) some strictly positive mass on the negative half axis.
iii) The gamma distribution is not easy to calculate. However if 2 IIN then
2Z
2
2
. The latter distribution can be found in statistical tables.
Let us consider the compound Poisson case. Then
=
4
3
2
2
3
, =
2
_
3
2
2
=
2
2
3
, k =
2
2
_
3
2
3
=
_
2
2
2
3
_
.
Example 1.6 (continued). A simple calculation shows that
3
= 27. Thus
=
80 27
729
= 2.96296, =
2 3
27
= 0.222222, k = 20
_
1
2 9
27
_
= 6.66667.
Figure 1.2 shows the densities of S (solid line) and of its translated gamma approx-
imation (dashed line).
Now 2 = 5.92592 6. Let
Z
2
6
.
II P[S > p] II P[Z > p 6.66667] II P[
Z > 0.444444(p 6.66667)] = 0.05
and thus 0.444444(p 6.66667) = 12.59 or p = 34.9942. For the 1% case we obtain
0.444444(p 6.66667) = 16.81 or p = 44.4892. Because the coecient of skewness
1. RISK MODELS 17
10 20 30 40 50
0.01
0.02
0.03
0.04
0.05
0.06
Figure 1.2: Densities of S and its translated gamma approximation
= 1.1619 is not small it is not surprising that the premium becomes larger than the
normal approximation premium especially if we are in the tail of the distribution.
In any case the approximations of the quantiles are better than for the normal
approximation. Recall that the exact values are 33.94 and 42.99 respectively. The
approximation puts more weight in the tail of the distribution. In the domain where
we did our computations the true value was overestimated. Note that for smaller
quantiles the true value eventually will be underestimated.
1.9.3. The Edgeworth Approximation
Consider the random variable
Z =
S II E[S]
_
Var[S]
.
The Taylor expansion of log M
Z
(r) around r = 0 has the form
log M
Z
(r) = a
0
+a
1
r +a
2
r
2
2
+a
3
r
3
6
+a
4
r
4
24
+
where
a
k
=
d
k
log M
Z
(r)
dr
k
r=0
.
18 1. RISK MODELS
We already know that a
0
= 0, a
1
= II E[Z] = 0, a
2
= Var[Z] = 1 and a
3
= II E[(Z
II E[Z])
3
] = . For a
4
a simple calculation shows that
a
4
= II E[Z
4
] 4II E[Z
3
]II E[Z] 3II E[Z
2
]
2
+ 12II E[Z
2
]II E[Z]
2
6II E[Z]
4
= II E[Z
4
] 3
=
II E[(S II E[S])
4
]
Var[S]
2
3 .
We truncate the Taylor series after the term involving r
4
. The moment generating
function of Z can be written as
M
Z
(r) e
r
2
/2
e
a
3
r
3
/6+a
4
r
4
/24
e
r
2
/2
_
1 +a
3
r
3
6
+a
4
r
4
24
+a
2
3
r
6
72
_
.
The inverse of expr
2
/2 is easily found to be the normal distribution function (x).
For the other terms we derive
r
n
e
r
2
/2
=
_
(e
rx
)
(n)
t
(x) dx = (1)
n
_
e
rx
(n+1)
(x) dx .
Thus the inverse of r
n
e
r
2
/2
is (1)
n
times the n-th derivative of . The approxima-
tion yields
II P[Z z] (z)
a
3
6
(3)
(z) +
a
4
24
(4)
(z) +
a
2
3
72
(6)
(z) .
It can be computed that
(3)
(z) =
1
2
(z
2
1)e
z
2
/2
,
(4)
(z) =
1
2
(z
3
+ 3z)e
z
2
/2
,
(6)
(z) =
1
2
(z
5
+ 10z
3
15z)e
z
2
/2
.
This approximation is called the Edgeworth approximation.
Consider now the compound Poisson case. One can see directly that a
k
=
k
(
2
)
k/2
. Hence
II P[S x] = II P
_
Z
x
2
_
_
x
2
_
3
6
_
3
2
(3)
_
x
2
_
+
4
24
2
2
(4)
_
x
2
_
+
2
3
72
3
2
(6)
_
x
2
_
.
1. RISK MODELS 19
10 20 30 40 50
0.01
0.02
0.03
0.04
0.05
0.06
Figure 1.3: Densities of S and its Edgeworth approximation
Example 1.7. We cannot consider Example 1.6 because
4
= . Let us therefore
consider the compound Poisson distribution with = 20 and Pa(5,4) distributed
claim sizes. The we nd = 1,
2
= 8/3,
3
= 16 and
4
= 256. Figure 1.3
gives us the densities of the exact distribution (solid line) and of its Edgeworth
approximation (dashed line). The shape is approximated quite well close to the
mean value 20. The tail is overestimated a little bit. Let us compare the 5% and
1% quantiles. Numerically we nd 33.64 and 42.99, respectively. The exact values
are 33.77 and 41.63, respectively. We see that the quantiles are approximated quite
well. It is not surprising that the 5% quantile is approximated better. The tail of
the approximation decays like x
5
e
cx
2
for some constant c while the exact tail decays
like x
5
which is much slower.
1.9.4. The Normal Power Approximation
The idea of the normal power approximation is to approximate
Z =
S II E[S]
_
Var[S]
Z +
S
6
(
Z
2
1) ,
where
Z is a standard normal variable and
S
is the coecient of skewness of S.
The mean value of the approximation is 0. But the variance becomes 1 +
2
S
/18 and
the skewness is
S
+
3
S
/27. If
S
is not too large the approximation meets the rst
three moments of Z quite well. Let us now consider the inequality
Z +
S
6
(
Z
2
1) z .
20 1. RISK MODELS
0 10 20 30 40 50
0.00
0.01
0.02
0.03
0.04
0.05
0.06
Figure 1.4: Densities of S and its normal power approximation
This is equivalent to
_
Z +
3
S
_
2
1 +
6z
S
+
9
2
S
.
We next consider positive
Z +3/
S
only, because
S
should not be too large. Thus
Z +
3
1 +
6z
S
+
9
2
S
.
We therefore get the approximation
II P[S x]
_
1 +
6(x II E[S])
S
_
Var[S]
+
9
2
S
S
_
.
In the compound Poisson case the approximation reads
II P[S x]
_
1 +
6(x )
2
3
+
9
3
2
2
3
3
_
3
2
3
_
.
Example 1.6 (continued). Plugging in the parameters we nd the approximation
II P[S x]
_
_
1 +
6(x 20)3
27
+
9 20 3
3
27
2
3
20 3
3
27
_
=
_
_
2x 17
3
_
20
3
_
.
1. RISK MODELS 21
The approximation is only valid if x 8.5. The density of S (solid line) and of the
normal power approximation (dashed line) are given in Figure 1.4. The density of
the approximation has a pole at x = 8.5. One can see that the approximation does
not t the tail as well as the translated gamma approximation. We get a clearer
picture if we calculate the 5% quantile which is 35.2999. The 1% quantile is 44.6369.
The approximation overestimates the true values by about 4%. As we also can see
by comparision with Figure 1.2 the translated gamma approximation ts the tail
much better than the normal power approximation.
1.10. Premium Calculation Principles
Let the annual premium for a certain risk be and denote the losses in year i by S
i
and assume that they are iid.. The company has a certain initial capital w. Then
the capital of the company after year i is
X
i
= w +i
i
j=1
S
j
.
The stochastic process X
i
is called a random walk. From Lemma E.1 we conclude
that
i) if < II E[S
1
] then X
i
converges to a.s..
ii) if = II E[S
1
] then a.s.
lim
i
X
i
= lim
i
X
i
= .
iii) if > II E[S
1
] then X
i
converges to a.s. and there is a strictly positive proba-
bility that X
i
0 for all i IIN.
As an insurer we only have a chance to survive if > II E[S]. Therefore the latter
must hold for any reasonable premium calculation principle. Another argument will
be given in Chapter 2.
1.10.1. The Expected Value Principle
The most popular premium calculation principle is
= (1 +)II E[S]
where is a strictly positive parameter called safety loading. The premium is very
easy to calculate. One only has to estimate the quantity II E[S]. A disadvantage is
that the principle is not sensible to heavy tailed distributions.
22 1. RISK MODELS
1.10.2. The Variance Principle
A way of making the premium principle more sensible to higher risks is to use the
variance for determining the risk loading, i.e.
= II E[S] +Var[S]
for a strictly positive parameter .
1.10.3. The Standard Deviation Principle
A principle similar to the variance principle is
= II E[S] +
_
Var[S]
for a strictly positive parameter . Observe that the loss can be written as
S =
_
Var[S]
_
S II E[S]
_
Var[S]
_
.
The standardized loss is therefore the risk loading parameter minus a random vari-
able with mean 0 and variance 1.
1.10.4. The Modied Variance Principle
A problem with the variance principle is that changing the monetary unit changes
the security loading. A possibility to change this is to use
= II E[S] +
Var[S]
II E[S]
for some > 0.
1.10.5. The Principle of Zero Utility
This premium principle will be discussed in detail in Chapter 2. Denote by w the
initial capital of the insurance company. The worst thing that may happen for the
company is a very high accumulated sum of claims. Therefore high losses should be
weighted stronger than small losses. Hence the company chooses a utility function
v, which should have the following properties:
v(0) = 0,
1. RISK MODELS 23
v(x) is strictly increasing,
v(x) is strictly concave.
The rst property is for convenience only, the second one means that less losses are
preferred and the last condition gives stronger weights for higher losses.
The premium for the next year is computed as the value such that
v(w) = II E[v(w +p S)] . (1.3)
This means, the expected utility is the same whether the insurance contract is taken
or not.
Lemma 1.8.
i) If a solution to (1.3) exists, then it is unique.
ii) If II P[S = II E[S]] < 1 then > II E[S].
iii) The premium is independent of w for all loss distributions if and only if v(x) =
A(1 e
x
) for some A > 0 and > 0.
Proof. i) Let
1
>
2
be two solutions to (1.3). Then because v(x) is strictly
increasing
v(w) = II E[v(w +
1
S)] > II E[v(w +
2
S)] = v(w)
which is a contradiction.
ii) By Jensens inequality
v(w) = II E[v(w + S)] < v(w + II E[S])
and thus because v(x) is strictly increasing we get w + II E[S] > w.
iii) A simple calculation shows that the premium is independent of w if v(x) =
A(1 e
x
). Assume now that the premium is independent of w. Let II P[S = 1] =
1 II P[S = 0] = q and denote by (q) be the corresponding premium. Then
qv(w +(q) 1) + (1 q)v(w +(q)) = v(w) . (1.4)
Note that as a concave function v(x) is dierentiable almost everywhere. Taking
the derivative of (1.4) (at points where it is allowed) with respect to w yields
qv
t
(w +(q) 1) + (1 q)v
t
(w +(q)) = v
t
(w) . (1.5)
24 1. RISK MODELS
Varying q the function (q) takes all values in (0, 1). Changing w and q in such a
way that w+(q) remains constant shows that v
t
(w) is continuous. Thus v(w) must
be continuously dierentiable everywhere.
The derivative of (1.4) with respect to q is
v(w+(q)1)v(w+(q))+
t
(q)[qv
t
(w+(q)1)+(1q)v
t
(w+(q))] = 0 . (1.6)
The implicit function theorem gives that (q) is indeed dierentiable. Plugging in
(1.5) into (1.6) yields
v(w +(q) 1) v(w +(q)) +
t
(q)v
t
(w) = 0 . (1.7)
Note that
t
(q) > 0 follows immediately. The derivative of (1.7) with respect to w
is
v
t
(w +(q) 1) v
t
(w +(q)) +
t
(q)v
tt
(w) = 0
showing that v(w) is twice continuously dierentiable. From the derivative of (1.7)
with respect to q it follows that
t
(q)[v
t
(w+(q) 1) v
t
(w+(q))] +
tt
(q)v
t
(w) =
t
(q)
2
v
tt
(w) +
tt
(q)v
t
(w) = 0 .
Thus
tt
(q) 0 and
v
tt
(w)
v
t
(w)
=
tt
(q)
t
(q)
2
.
The left-hand side is independent of q and the right-hand side is independent of w.
Thus it must be constant. Hence
v
tt
(w)
v
t
(w)
=
for some 0. = 0 would not lead to a strictly concave function. Solving the
dierential equation shows the assertion.
1.10.6. The Mean Value Principle
In contrast to the principle of zero utility it would be desirable to have a premium
principle which is independent of the initial capital. The insurance company values
its losses and compares it to the loss of the customer who has to pay its premium.
Because high losses are less desirable they need to have a higher value. Similarly to
the utility function we choose a value function v with properties
1. RISK MODELS 25
v(0) = 0,
v(x) is strictly increasing and
v(x) is strictly convex.
The premium is determined by the equation
v() = II E[v(S)]
or equivalently
= v
1
(II E[v(S)]) .
Again, it follows from Jensens inequality that > II E[S] provided II P[S = II E[S]] < 1.
1.10.7. The Exponential Principle
Let > 0 be a parameter. Choose the utility function v(x) = 1 e
x
. The
premium can be obtained
1 e
w
= II E
_
1 e
(w+pS)
= 1 e
w
e
p
II E
_
e
S
. (1.8)
The same premium formula can be obtained by using the mean value principle with
value function v(x) = e
x
1. Note that the principle only can be applied if the
moment generating function of S exists at the point .
How does the premium depend on the parameter ? In order to answer this
question the following lemma is proved.
Lemma 1.9. For a random variable X with moment generating function M
X
(r)
the function log M
X
(r) is convex. If X is not deterministic then log M
X
(r) is strictly
convex.
Proof. Let F(x) denote the distribution function of X. The second derivative of
log M
X
(r) is
(log M
X
(r))
tt
=
M
tt
X
(r)
M
X
(r)
_
M
t
X
(r)
M
X
(r)
_
2
.
Consider
M
tt
X
(r)
M
X
(r)
=
d
2
dr
2
_
e
rx
dF(x)
M
X
(r)
=
_
x
2
e
rx
dF(x)
M
X
(r)
26 1. RISK MODELS
and
M
t
X
(r)
M
X
(r)
=
_
xe
rx
dF(x)
M
X
(r)
.
Because
_
e
rx
dF(x)
M
X
(r)
=
M
X
(r)
M
X
(r)
= 1
it follows that F
r
(x) = (M
X
(r))
1
_
x
e
ry
dF(y) is a distribution function. Let Z
be a random variable with distribution function F
r
. Then
(log M
X
(r))
tt
= Var[Z] 0 .
The variance is strictly larger than 0 if Z is not deterministic. But the latter is
equivalent to X is not deterministic.
The question how the premium depends on can now be answered.
Lemma 1.10. Let () be the premium determined by the exponential principle.
Assume that M
tt
S
() < . The function () is strictly increasing provided S is not
deterministic.
Proof. Take the derivative of (1.8)
t
() =
M
t
S
()
M
S
()
log(M
S
())
2
=
1
_
M
t
S
()
M
S
()
()
_
.
Furthermore
d
d
(
2
t
()) =
M
t
S
()
M
S
()
() +
_
M
tt
S
()
M
S
()
_
M
t
S
()
M
S
()
_
2
_
t
()
=
_
M
tt
S
()
M
S
()
_
M
t
S
()
M
S
()
_
2
_
.
From Lemma 1.9 it follows that
d
d
(
2
t
()) 0 i 0 .
Thus the function
2
t
() has a minimum in 0. Because its value at 0 vanishes it fol-
lows that
t
() > 0.
t
(0) =
1
2
Var[S] > 0 can be veried directly with LHospitals
rule.
1. RISK MODELS 27
1.10.8. The Esscher Principle
A simple idea for pricing is to use a distribution
F(x) which is related to the dis-
tribution function of the risk F(x), but does give more weight to larger losses. If
exponential moments exist, a natural choice would be to consider the exponential
class
F
(x) = (M
S
())
1
_
x
e
ry
dF
S
(y) .
This yields the Esscher premium principle
=
II E[Se
S
]
M
S
()
,
where > 0. In order that the principle can be used we have to show that > II E[S]
if Var[S] > 0. Indeed, = (log M
S
(r))
t
[
r=
and log M
S
(r) is a strictly convex
function by Lemma 1.9. Thus (log M
S
(r))
t
is strictly increasing, in particular >
(log M
S
(r))
t
[
r=0
= II E[S].
1.10.9. The Proportional Hazard Transform Principle
A disadvantage of the Esscher principle is, that M
S
() has to exist. We therefore
look for a possibility to increase the weight for large claims, such that no exponential
moments need to exist. A possibility is to change the distribution function F
S
(x) to
F
r
(x) = 1 (1 F
S
(x))
r
, where r (0, 1). Then F
r
(x) < F
S
(x) for all x such that
F
S
(x) (0, 1). The proportional hazard transform principle is
=
_
0
(1 F
S
(x))
r
dx
_
0
1 (1 F
S
(x))
r
dx .
Note that clearly > II E[S]. Because dF
r
(x) = r(1 F
S
(x))
r1
dF
S
(x) it follows
that the distribution F
r
(x) and F
S
(x) are equivalent. A sucient condition for to
be nite would be that II E[maxS, 0
n+1,2j+1
,
v(
n+1,2j+1
) >
1
2
(v(
j,n
) +v(
n,j+1
)) >
1
2
(
n,j
+
n,j+1
) =
n+1,2j+1
.
Let now be arbitrary and let j
n
such that
n,jn
<
n,jn+1
,
n
=
n,jn
. Then
v() v(
n
) >
n
.
Because n is arbitrary we have v() . This implies that u(y) is concave. Let
now <
1
2
. Then
u((1 )x +z) = u((1 2)x + 2
1
2
(x +z)) (1 2)u(x) + 2u(
1
2
(x +z))
> (1 2)u(x) + 2
1
2
(u(x) +u(z)) = (1 )u(x) +u(z) .
A similar argument applies if >
1
2
. The converse statement is left as an exercise.
Next we would like to have a measure, how much an agent likes the risk. Assume
that u(y) is twice dierentiable. Intuitively, the more concave a utility function is,
the more it is sensible to risk. So a risk averse agent will have a utility function
with u
tt
(y) large. But u
tt
(y) is not a good measure for this purpose. Let u(y) =
a+bu(y). Then u
tt
(y) = bu
tt
(y) would give a dierent risk aversion, even though
it leads to the same decisions. In order to get rid of b dene the risk aversion
function
a(y) =
u
tt
(y)
u
t
(y)
=
d
dy
ln u
t
(y) .
Intuitively, we would assume that a wealthy person can take more risk than a poor
person. This would mean that a(y) should be decreasing. For our examples we have
that (2.1a) and (2.1e) give increasing risk aversion, (2.1c) and (2.1d) give decreasing
risk aversion, and (2.1b) gives constant risk aversion.
32 2. UTILITY THEORY
2.2. The Zero Utility Premium
Let us consider the problem that an agent has to decide whether to insure his risk
or not. Let (, r(x)) be an insurance treaty and assume that it is a non-trivial
insurance, i.e. Var[r(X)] > 0. If the agent takes the insurance he will have the nal
wealth ws(X), if he does not take the insurance his nal wealth will be wX.
Considering his expected utility, the agent will take the insurance if and only if
II E[u(w s(X))] II E[u(w X)] .
Because insurance basically reduces the risk we assume here that insurance is taken
if the expected utility is the same. The left hand side of the above inequality is
a continuous and strictly decreasing function of . The largest value that can be
attained is for = 0, II E[u(w s(X))] > II E[u(w X)], the lowest possible value is
dicult to obtain, because it depends on the choice of I. Let us assume that the
lower bound is smaller or equal to II E[u(w X)]. Then there exists a unique value
r
(w) such that
II E[u(w
r
s(X))] = II E[u(w X)] .
The premium
r
is the largest premium the agent is willing to pay for the insurance
treaty r(x). We call
r
the zero utility premium. In many situations the zero
utility premium
r
will be larger then the fair premium II E[r(X)].
Theorem 2.2. If both r(x) and s(x) are increasing and Var[r(X)] > 0 then
r
(w) > II E[r(X)].
Proof. From the denition of a strictly concave function it follows that the func-
tion v(y), x v(x) = u(w x) is a strictly concave function. Dene the func-
tions g
1
(x) = s(x) + II E[r(X)] and g
2
(x) = x. These two functions are increasing.
The function g
2
(x) g
1
(x) = r(x) II E[r(X)] is increasing. Thus there exists x
0
such that g
1
(x) g
2
(x) for x < x
0
and g
1
(x) g
2
(x) for x > x
0
. Note that
II E[g
2
(x) g
1
(x)] = 0. Thus the conditions of Corollary G.7 are fullled and
II E[u(w II E[r(X)] s(X))] = II E[v(g
1
(X))] > II E[v(g
2
(X))]
= II E[u(w X)] = II E[u(w
r
s(X))] .
Because II E[u(w s(X))] is a strictly decreasing function it follows that
r
> II E[r(X)].
Assume now that u is twice continuously dierentiable. Then we nd the follow-
ing result.
2. UTILITY THEORY 33
Theorem 2.3. Let r(x) = x. If the risk aversion function is decreasing (increas-
ing), then the zero utility premium (w) is decreasing (increasing).
Proof. Consider the function v(y) = u
t
(u
1
(y)). Taking the derivative yields
v
t
(y) =
u
tt
(u
1
(y))
u
t
(u
1
(y))
= a(u
1
(y)) .
Because y u
1
(y) is increasing, this gives that v
t
(y) is increasing (decreasing).
Thus v(y) is convex (concave).
We have
(w) = w u
1
(II E[u(w X)]) and
t
(w) = 1
II E[u
t
(w X)]
u
t
(u
1
(II E[u(w X)]))
.
Let now a(y) be decreasing (increasing). By Jensens inequality we have
u
t
(u
1
(II E[u(w X)])) () II E[u
t
(u
1
(u(w X)))] = II E[u
t
(w X)]
and therefore
t
(w) () 0.
It is not always desirable that the decision depends on the initial wealth. We
have the following result.
Lemma 2.4. Let u(x) be a utility function. Then the following are equivalent:
i) u(x) is exponential, i.e. u(x) = Ae
cx
+B for some A, c > 0, B IR.
ii) For all losses X and all compensation functions r(x), the zero utility premium
r
does not depend on the initial wealth w.
iii) For all losses X and for full reinsurance r(x) = x, the zero utility premium
r
does not depend on the initial wealth w.
Proof. This follows directly from Lemma 1.8.
2.3. Optimal Insurance
Let us now consider an agent carrying several risks X
1
, X
2
, . . . , X
n
. The total risk is
then X = X
1
+ +X
n
. We consider here a compensation function r(X
1
, . . . , X
n
)
and a self-insurance function s(X
1
, . . . , X
n
) = X r(X
1
, . . . , X
n
). We say an
insurance treaty is global if r(X
1
, . . . , X
n
) depends on X only, or equivalently,
s(X
1
, . . . , X
n
) depends on X only. Otherwise, we call an insurance treaty local. We
next prove that, in some sense, the global treaties are optimal for the agent.
34 2. UTILITY THEORY
Theorem 2.5. (Pesonen-Ohlin) If the premium depends on the pure net pre-
mium only, then to every local insurance treaty there exists a global treaty with the
same premium but a higher expected utility.
Proof. Let (, r(x
1
, . . . , x
n
)) be a local treaty. Consider the compensation func-
tion R(x) = II E[r(X
1
, . . . , X
n
) [ X = x]. If the premium depends on the pure net
premium only, then clearly R(X) is sold for the same premium, because
II E[R(X)] = II E[II E[r(X
1
, . . . , X
n
) [ X]] = II E[r(X
1
, . . . , X
n
)] .
By Jensens inequality
II E[u(w s(X
1
, . . . , X
n
))] = II E[II E[u(w s(X
1
, . . . , X
n
)) [ X]]
< II E[u(II E[w s(X
1
, . . . , X
n
) [ X])] = II E[u(w (X R(X)))] .
The strict inequality follows from the fact that s(X
1
, . . . , X
n
) ,= S(X), because
otherwise the insurance treaty would be global, and from the strict concavity.
Remark. Note that the result does not say that the constructed global treaty is
a good treaty. It only says, that for the investigation of good treaties we can restrict
to global treaties.
Let us now nd the optimal insurance treaty for the insured.
Theorem 2.6. (Arrow-Ohlin) Assume the premium depends on the pure net
premium only and let (, r(x)) be an insurance treaty with II E[r(X)] > 0. Then there
exists a unique deductible b such that the excess of loss insurance r
b
(x) = (xb)
+
has the same premium. Moreover, if II P[r(X) ,= r
b
(X)] > 0 then the treaty (, r
b
(x))
yields a higher utility.
Proof. The function (x b)
+
is continuous and decreasing in b. We have II E[(X
0)
+
] II E[r(X)] > 0 = lim
b
II E[(X b)
+
]. Thus there exists a unique b such
that II E[(X b)
+
] = II E[r(X)]. Thus r
b
(x) yields the same premium. Assume now
II P[r(X) ,= r
b
(X)] > 0. Let s
b
(x) = x r
b
(x). For y < b we have
II P[s
b
(X) y] = II P[X y] II P[s(X) y] ,
and for y > b
II P[s
b
(X) y] = 1 II P[s(X) y] .
2. UTILITY THEORY 35
Note that v(x) = u(w x) is a strictly concave function. Thus Ohlins lemma
yields
II E[u(w s
b
(X))] = II E[v(s
b
(X))] > II E[v(s(X))] = II E[u(w s(X))] .
r
does not depend on the initial wealth w.
Proof. The result follows analogously to the proof of Proposition 2.4.
2.5. Pareto-Optimal Risk Exchanges
Consider now a market with n agents, i.e. insured, insurance companies, reinsurers.
In the market there are risks X = (X
1
, . . . , X
n
)
: IR
n
IR
n
,
such that
n
i=1
f
i
(X) =
n
i=1
X
i
. (2.3)
The latter condition assures that the whole risk is covered. A risk exchange means
that agent i covers f
i
(X). This is a generalisation of the insurance market we had
considered before.
Now each of the agents has an initial wealth w
i
and a utility function u
i
(x).
Under the risk exchange f the expected terminal utility of agent i becomes
V
i
(f) = II E[u
i
(w
i
f
i
(X))] .
As we saw before, the interests of the agents will contradict. Thus it will in general
not be possible to nd f such that each agent gets an optimal utility. However, an
agent may agree to a risk exchange f if V
i
(f) II E[u
i
(w
i
X
i
)]. Hence, which risk
exchange will be chosen depends upon negotiations. However, the agents will agree
not to discuss certain risk exchanges.
Denition 2.9. A risk exchange f is called Pareto-optimal if for any risk ex-
change
f with
V
i
(f) V
i
(
f) i it follows that V
i
(f) = V
i
(
f) i .
For a risk exchange f that is not Pareto optimal we would have a risk exchange
f
that is at least as good for all agents, but better for at least one of them. Of course,
it is not excluded that V
i
(f) < V
i
(
f) for some i. If the latter is the case then there
must be j such that V
j
(f) > V
j
(
f).
For the negotiations, the agents can in principle restrict to Pareto-optimal risk
exchanges. Of course, in reality, the agents do not know the utility functions of
the other agents. So the concept of Pareto optimality is more a theoretical way to
consider the market.
It seems quite dicult to decide whether a certain risk exchange is Pareto-optimal
or not. It turns out that a simple criteria does the job.
38 2. UTILITY THEORY
Theorem 2.10. (Borch/du Mouchel) Assume that the utility functions u
i
(x)
are dierentiable. A risk exchange f is Pareto-optimal if and only if there exist
numbers
i
such that (almost surely)
u
t
i
(w
i
f
i
(X)) =
i
u
t
1
(w
1
f
1
(X)) . (2.4)
Remark. Note that necessarily
1
= 1 and
i
> 0 because u
t
i
(x) > 0. Note also
that we could make the comparisons with agent j instead of agent 1. Then we just
need to divide the
i
by
j
, for example
1
would be replaced by
1
j
.
Proof. Assume that
i
exists such that (2.4) is fullled. Let
f be a risk exchange
such that V
i
(
f) V
i
(f) for all i. By concavity
u
i
(w
i
f
i
(X)) u
i
(w
i
f
i
(X)) +u
t
i
(w
i
f
i
(X))(f
i
(X)
f
i
(X)) .
This yields
u
i
(w
i
f
i
(X)) u
i
(w
i
f
i
(X))
i
u
t
1
(w
1
f
1
(X))(f
i
(X)
f
i
(X)) .
Using
n
i=1
f
i
(X) =
n
i=1
X
i
=
n
i=1
f
i
(X) we nd by summing over all i,
n
i=1
u
i
(w
i
f
i
(X)) u
i
(w
i
f
i
(X))
i
0 .
Taking expectations yields
n
i=1
V
i
(
f) V
i
(f)
i
0 .
From the assumption V
i
(
f) V
i
(f) it follows that V
i
(
f) = V
i
(f). Thus f is
Pareto-optimal.
Assume now that (2.4) does not hold. By renumbering, we can assume that
there is no constant
2
such that (2.4) holds for i = 2. Let
=
II E[u
t
1
(w
1
f
1
(X))u
t
2
(w
2
f
2
(X))]
II E[(u
t
1
(w
1
f
1
(X)))
2
]
and W = u
t
2
(w
2
f
2
(X))u
t
1
(w
1
f
1
(X)). We have then II E[Wu
t
1
(w
1
f
1
(X))] = 0
and by the assumption II E[W
2
] > 0. Let now > 0, =
1
2
II E[W
2
]/II E[u
t
2
(w
2
2. UTILITY THEORY 39
f
2
(X))] > 0 and
f
1
(X) = f
1
(X) ( W),
f
2
(X) = f
2
(X) +( W),
f
i
(X) =
f
i
(X) for all i > 2. Then
f is a risk exchange. Consider now
V
1
(
f) V
1
(f)
II E[u
t
1
(w
1
f
1
(X))]
= II E
_
( W)
_
u
1
(w
1
f
1
(X)) u
1
(w
1
f
1
(X))
( W)
u
t
1
(w
1
f
1
(X))
__
,
where we used II E[Wu
t
1
(w
1
f
1
(X))] = 0. The right-hand side tends to zero as
tends to zero. Thus V
1
(
f) > V
1
(f) for small enough. We also have
V
2
(
f) V
2
(f)
II E[u
t
2
(w
2
f
2
(X))(W )]
= II E
_
(W )
_
u
2
(w
2
f
2
(X)) u
2
(w
2
f
2
(X))
(W )
u
t
2
(w
2
f
2
(X))
__
,
Also here the right-hand side tends to zero. We have
II E[u
t
2
(w
2
f
2
(X))(W )] = II E[(W +u
t
1
(w
1
f
1
(X)))W] II E[u
t
2
(w
2
f
2
(X))]
= II E[W
2
] II E[u
t
2
(w
2
f
2
(X))] =
1
2
II E[W
2
] > 0 .
Thus also V
2
(
f) > V
2
(f) for small enough. This shows that f is not Pareto-
optimal.
In order to nd the Pareto-optimal risk exchanges one has to solve equation (2.4)
subject to the constraint (2.3).
If a risk exchange is chosen the quantity f
i
(0, . . . , 0) is the amount agent i has
to pay (obtains if negative) if no losses occur. Thus f
i
(0, . . . , 0) must be interpreted
as a premium.
Note that for Pareto-optimality only the support of the distribution of X, not
the distribution itself, does has an inuence. Hence the Pareto-optimal solution can
be found without investigating the risk distribution, or the dependencies. This also
shows that any Pareto optimal solution will depend on
n
i=1
X
i
only.
We now solve the problem explicitly. Because u
t
i
(x) is strictly decreasing, it is
invertible and its inverse (u
t
i
)
1
(y) is strictly decreasing. Moreover, if u
t
i
(x) is con-
tinuous (as it is under the conditions of Theorem 2.10), then (u
t
i
)
1
(y) is continuous
too. Choose
i
> 0,
1
= 1. Then (2.4) yields
w
i
f
i
(X) = (u
t
i
)
1
(
i
u
t
1
(w
1
f
1
(X))) .
40 2. UTILITY THEORY
Summing over i gives by use of (2.3)
n
i=1
w
i
i=1
X
i
=
n
i=1
(u
t
i
)
1
(
i
u
t
1
(w
1
f
1
(X))) .
The function
g(y) =
n
i=1
(u
t
i
)
1
(
i
u
t
1
(y))
is then strictly increasing and therefore invertible. If all the u
t
i
(x) are continuous,
then also g(y) is continuous and therefore g
1
(x) is continuous too. This yields
f
1
(X) = w
1
g
1
_
n
i=1
w
i
i=1
X
i
_
.
For i arbitrary we then obtain
f
i
(X) = w
i
(u
t
i
)
1
_
i
u
t
1
_
g
1
_
n
i=1
w
i
i=1
X
i
___
.
Thus we have found the following result.
Theorem 2.11. A Pareto-optimal risk exchange is a pool, that is each individual
agent contributes a share that depends on the total loss of the group only. Moreover,
each agent must cover a genuine part of any increase of the total loss of the group.
If all utility functions are dierentiable, then the individual shares are continuous
functions of the total loss.
Remark. Consider the situation insured-insurer, i.e. n = 2 and a risk of the form
(X, 0)
i=1
i
be the mean number of years with an accident for a typical driver. Then
Z =
10
20
i=1
(
i
)
2
(1 )
= 49.1631 .
Under the hypothesis that II E[
i
] is the same for all drivers the statistic Z is approx-
imately
2
19
-distributed. But
II P[
2
19
> 49] 0.0002 .
42 3. CREDIBILITY THEORY
Thus it is very unlikely that each driver has the same probability of having at least
one accident in a year.
For the insurance company it is preferable to attract good risks and to get
rid of the poor risks. Thus they would like to charge premia according to their
experience with a particular customer. This procedure is called experience rating
or credibility. Let us denote by Y
ij
the losses of the i-th risk in year j, where
i = 1, 2, . . . , n and j = 1, 2, . . . , m. Denote the mean losses of the i-th risk by
Y
i
:=
1
m
m
j=1
Y
ij
.
We make the following assumptions:
i) There exists a parameter
i
belonging to the risk i. The parameters (
i
: 1
i n) are iid..
ii) The vectors ((Y
i1
, Y
i2
, . . . , Y
im
,
i
) : 1 i n) are iid..
iii) For xed i, given
i
the random variables Y
i1
, Y
i2
, . . . Y
im
are conditionally iid..
For instance
i
can be the parameters of a family of distributions, or more generally,
the distribution function of a claim of the i-th risk.
Denote by m() = II E[Y
ij
[
i
= ] the conditional expected mean of an aggregate
loss given
i
= , by = II E[m(
i
)] the overall mean of the aggregate losses and
by s
2
() = Var[Y
ij
[
i
= ] the conditional variance of Y
ij
given
i
= . Now
instead of using the expected mean of a claim of a typical risk for the premium
calculations one should use m(
i
) instead. Unfortunately we do not know
i
. But
Y
i
is an estimator for m(
i
). But it is not a good idea either to use
Y
i
for the
premium calculation. Assume that somebody who is insured for the rst year has
a big claim. His premium for the next year would become larger than his claim.
He would not be able to take insurance furthermore. Thus we have to nd other
methods to estimate m(
i
).
3.2. Bayesian Credibility
Suppose we know the distribution of
i
and the conditional distribution of Y
ij
given
i
. We now want to nd the best estimate M
0
of m(
i
) in the sense that
II E[(M m(
i
))
2
] II E[(M
0
m(
i
))
2
]
3. CREDIBILITY THEORY 43
for all measurable functions M = M(Y
i1
, Y
i2
, . . . , Y
im
). Let M
t
0
= II E[m(
i
) [
Y
i1
, Y
i2
, . . . , Y
im
]. Then
II E[(M m(
i
))
2
] = II E[(M M
t
0
) + (M
t
0
m(
i
))
2
]
= II E[(M M
t
0
)
2
] + II E[(M
t
0
m(
i
))
2
] + 2II E[(M M
t
0
)(M
t
0
m(
i
))] .
We have then
II E[(M M
t
0
)(M
t
0
m(
i
))] = II E[II E[(M M
t
0
)(M
t
0
m(
i
)) [ Y
i1
, Y
i2
, . . . , Y
im
]]
= II E[(M M
t
0
)(M
t
0
II E[m(
i
) [ Y
i1
, Y
i2
, . . . , Y
im
])]
= 0 .
Thus
II E[(M m(
i
))
2
] = II E[(M M
t
0
)
2
] + II E[(M
t
0
m(
i
))
2
] .
The second term is independent of the choice of M. That means that the best
estimate is M
0
= M
t
0
. We call M
0
the Bayesian credibility estimator. In the sequel
we will consider the two most often used Bayesian credibility models.
3.2.1. The Poisson-Gamma Model
We assume that Y
ij
has a compound Poisson distribution with parameter
i
and an
individual claim size distribution which is the same for all the risks. It is therefore
sucient for the company to estimate
i
. In the sequel we assume that all claim
sizes are equal to 1, i.e. Y
ij
= N
ij
. Motivated by the compound negative binomial
model we assume that
i
(, ). Let
N
i
=
1
m
m
j=1
N
ij
44 3. CREDIBILITY THEORY
denote the mean number of claims of the i-th risk. We know the overall mean
II E[
i
] =
1
. For the best estimator for
i
we get
II E[
i
[ N
i1
, N
i2
, . . . , N
im
] =
_
0
j=1
_
N
ij
N
ij
!
e
()
1
e
d
_
0
m
j=1
_
N
ij
N
ij
!
e
()
1
e
d
=
_
0
m
N
i
+
e
(m+)
d
_
0
m
N
i
+1
e
(m+)
d
=
(m
N
i
+ + 1)(m+)
m
N
i
+
(m+)
m
N
i
++1
(m
N
i
+)
=
m
N
i
+
m+
=
m
m+
N
i
+
m+
_
_
=
m
m+
N
i
+
_
1
m
m+
_
II E[
i
] .
The best estimator is therefore of the form
Z
N
i
+ (1 Z)II E[
i
]
with the credibility factor
Z =
1
1 +
m
.
The credibility factor Z is increasing with the number of years m and converges to
1 as m . Thus the more observations we have the higher is the weight put to
the empirical mean
N
i
.
The only problem that remains is to estimate the parameters and . For an
insurance company this quite often is no problem because they have big data sets.
3.2.2. The Normal-Normal Model
For the next model we do not assume that the individual claims have the same
distribution for all risks. First we assume that each risk consists of a lot of sub-
risks such that the random variables Y
ij
are approximately normally distributed.
Moreover, the conditional variance is assumed to be the same for all risks, i.e. Y
ij
N(
i
,
2
). Next we assume that the parameter
i
N(,
2
) is also normally
distributed. Note that and
2
have to be chosen in such a way that II P[Y
ij
0] is
small.
Let us rst consider the joint density of Y
i1
, Y
i2
, . . . Y
im
,
i
(2)
(m+1)/2
m
exp
_
_
( )
2
2
2
+
m
j=1
(y
ij
)
2
2
2
__
.
3. CREDIBILITY THEORY 45
Because we will mainly be interested in the posterior distribution of
i
we should
write the exponent in the form ( )
2
. For abbreviation we write m y
i
=
m
j=1
y
ij
.
( )
2
2
2
+
m
j=1
(y
ij
)
2
2
2
=
2
_
1
2
2
+
m
2
2
_
2
_
2
2
+
m y
i
2
2
_
+
2
2
2
+
m
j=1
y
2
ij
2
2
.
The joint density can therefore be written as
C(y
i1
, y
i2
, y
im
) exp
_
_
_
_
_
2
+
m y
i
2
__
1
2
+
m
2
_
1
_
2
2
_
1
2
+
m
2
_
1
_
_
_
where C is a function of the data. Hence the posterior distribution of
i
is normally
distributed with mean
_
2
+
m
Y
i
2
__
1
2
+
m
2
_
1
=
2
+m
2
Y
i
2
+m
2
=
m
2
2
+m
2
Y
i
+
_
1
m
2
2
+m
2
_
.
Again the credibility premium is a linear combination Z
Y
i
+ (1 Z)II E[Y
ij
] of the
mean losses of the i-th risk and the overall mean . The credibility factor has a
similar form as in the Poisson-gamma model
Z =
1
1 +
2
2
m
.
3.2.3. Is the Credibility Premium Formula Always Linear?
From the previous considerations one could conjecture that the credibility premium
is always of the form Z
Y
i
+(1Z)II E[Y
ij
]. We try to nd a counterexample. Assume
that
i
takes the values 1 and 2 with equal probabilities
1
2
. Given
i
the random
variables Y
ij
Pois(
i
) are Poisson distributed. Then
II E[
i
[ Y
i1
, Y
i2
, . . . , Y
im
]
= II P[
i
= 1 [ Y
i1
, Y
i2
, . . . , Y
im
] + 2II P[
i
= 2 [ Y
i1
, Y
i2
, . . . , Y
im
]
= 2 II P[
i
= 1 [ Y
i1
, Y
i2
, . . . , Y
im
] = 2
1
2
m
j=1
1
Y
ij
!
e
1
1
2
_
m
j=1
1
Y
ij
!
e
1
+
m
j=1
2
Y
ij
Y
ij
!
e
2
_
= 2
1
1 + e
m
2
m
Y
i
.
It turns out that there is no possibility to get a linear formula.
46 3. CREDIBILITY THEORY
3.3. Empirical Bayes Credibility
In order to calculate the Bayes credibility estimator one needs to know the joint
distribution of m(
i
) and (Y
i1
, Y
i,2
, . . . , Y
im
). In practice one does not have distri-
butions but only data. It seems therefore not feasible to estimate joint distributions
if only one of the variables is observable. We therefore want to restrict now to linear
estimators, i.e. estimators of the form
M = a
i0
+a
i1
Y
i1
+a
i2
Y
i2
+ +a
im
Y
im
.
The best estimator is called linear Bayes estimator. It turns out that it is enough
to know the rst two moments of (m(
i
), Y
i
1). One then has to estimate these
quantities. We further do not know the mean values and the variances. In practice
we will have to estimate these quantities. We will therefore now proceed in two
steps: First we estimate the linear Bayes premium, and then we estimate the rst
two moments. The corresponding estimator is called empirical Bayes estimator.
3.3.1. The B uhlmann Model
In addition to the notation introduced before denote by
2
= II E[s
2
(
i
)] and by
v
2
= Var[m(
i
)]. Note that
2
is not the unconditional variance of Y
ij
. In fact
II E[Y
2
ij
] = II E[II E[Y
2
ij
[
i
]] = II E[s
2
(
i
) + (m(
i
))
2
] =
2
+ II E[(m(
i
))
2
] .
Thus the variance of Y
ij
is
Var[Y
ij
] =
2
+ II E[(m(
i
))
2
]
2
=
2
+v
2
. (3.1)
We consider credibility premia of the form
a
i0
+
m
j=1
a
ij
Y
ij
.
Which parameters a
ij
minimize the expected quadratic error
II E
__
a
i0
+
m
j=1
a
ij
Y
ij
m(
i
)
_
2
_
?
We rst dierentiate with respect to a
i0
.
1
2
d
da
i0
II E
__
a
i0
+
m
j=1
a
ij
Y
ij
m(
i
)
_
2
_
= II E
_
a
i0
+
m
j=1
a
ij
Y
ij
m(
i
)
_
= a
i0
+
_
m
j=1
a
ij
1
_
= 0 .
3. CREDIBILITY THEORY 47
It follows that
a
i0
=
_
1
m
j=1
a
ij
_
j=1
a
ij
Y
ij
_
= .
Let 1 k m and dierentiate with respect to a
ik
.
1
2
d
da
ik
II E
__
a
i0
+
m
j=1
a
ij
Y
ij
m(
i
)
_
2
_
= II E
_
Y
ik
_
a
i0
+
m
j=1
a
ij
Y
ij
m(
i
)
__
= a
i0
+
m
j=1
j=k
a
ij
II E[Y
ik
Y
ik
] +a
ik
II E[Y
2
ik
] II E[Y
ik
m(
i
)] = 0 .
We already know from (3.1) that II E[Y
2
ik
] =
2
+v
2
+
2
. For j ,= k we get
II E[Y
ik
Y
ij
] = II E[II E[Y
ik
Y
ij
[
i
]] = II E[(m(
i
))
2
] = v
2
+
2
.
And nally
II E[Y
ik
m(
i
)] = II E[II E[Y
ik
m(
i
) [
i
]] = II E[(m(
i
))
2
] = v
2
+
2
.
The equations to solve are
_
1
m
j=1
a
ij
_
2
+
m
j=1
j=k
a
ij
(v
2
+
2
) +a
ik
(
2
+v
2
+
2
) (v
2
+
2
)
= a
ik
_
1
m
j=1
a
ij
_
v
2
= 0 .
The right hand side of
2
a
ik
= v
2
_
1
m
j=1
a
ij
_
is independent of k. Thus
a
i1
= a
i2
= = a
im
=
v
2
2
+mv
2
. (3.2)
The credibility premium is
mv
2
2
+mv
2
Y
i
+
_
1
mv
2
2
+mv
2
_
= Z
Y
i
+ (1 Z) (3.3)
48 3. CREDIBILITY THEORY
where
Z =
1
1 +
2
v
2
m
.
The formula for the credibility premium does only depend on ,
2
, and v
2
. This
also were the only quantities we had assumed in our model. Hence the approach is
quite general. We do not need any other assumptions on the distribution of
i
.
In order to apply the result we need to estimate the parameters ,
2
, and v
2
.
In order to make the right decision we look for unbiased estimators.
i) : The natural estimator of is
=
1
nm
n
i=1
m
j=1
Y
ij
=
1
n
n
i=1
Y
i
. (3.4)
It is easy to see that is unbiased. It can be shown that is the best linear
unbiased estimator.
ii)
2
: For the estimation of s
2
(
i
) we would use the unbiased estimator
1
m1
m
j=1
(Y
ij
Y
i
)
2
.
Thus
2
=
1
n(m1)
n
i=1
m
j=1
(Y
ij
Y
i
)
2
(3.5)
is an unbiased estimator for
2
.
iii) v
2
:
Y
i
is an unbiased estimator of m(
i
). Therefore a natural estimate of v
2
would be
1
n 1
n
i=1
(
Y
i
)
2
.
3. CREDIBILITY THEORY 49
But is this estimator really unbiased?
1
n 1
n
i=1
II E[(
Y
i
)
2
] =
1
n 1
n
i=1
II E
__
n 1
n
Y
i
1
n
j,=i
Y
j
_
2
_
=
n
n 1
II E
__
n 1
n
Y
1
1
n
n
j=2
Y
j
_
2
_
=
n
n 1
II E
__
n 1
n
(
Y
1
)
1
n
n
j=2
(
Y
j
)
_
2
_
=
n
n 1
__
n 1
n
_
2
II E[(
Y
1
)
2
] +
n 1
n
2
II E[(
Y
1
)
2
]
_
= II E
_
(
Y
1
)
2
= II E
_
1
m
2
m
i=1
m
j=1
Y
1i
Y
1j
2
m
j=1
Y
1j
+
2
_
=
1
m
(
2
+v
2
+
2
) +
m1
m
(v
2
+
2
)
2
= v
2
+
2
m
.
Our proposal for an estimator turns out to be biased. We have to correct the
estimator and get
v
2
=
1
n 1
n
i=1
(
Y
i
)
2
1
m
2
. (3.6)
Example 3.1. A company has insured twenty similar collective risks over the last
ten years. Table 3.2 shows the annual losses of each the risks. Thus the estimators
are
=
1
20
20
i=1
Y
i
= 102.02 ,
2
=
1
180
20
i=1
10
j=1
(Y
ij
Y
i
)
2
= 473.196
and
v
2
=
1
19
20
i=1
(
Y
i
)
2
1
10
2
= 806.6 .
The credibility factor can now be computed
Z =
1
1 +
2
v
2
m
= 0.944585 .
Table 3.3 gives the credibility premia for next year for each of the risks. The data
had been simulated using = 100,
2
= 400 and v
2
= 900. The credibility model
50 3. CREDIBILITY THEORY
i 1 2 3 4 5 6 7 8 9 10
Y
i
1 67 71 50 56 64 69 77 94 56 48 65.2
2 80 82 109 61 89 113 149 91 127 108 100.9
3 125 118 101 135 120 117 101 101 100 122 114.0
4 96 144 152 124 94 132 155 143 94 113 124.7
5 176 161 153 191 139 157 192 151 175 147 164.2
6 89 129 95 88 131 72 91 122 79 56 95.2
7 70 116 102 129 81 69 75 120 95 93 95.0
8 22 33 48 20 56 25 43 53 48 64 41.2
9 121 106 55 103 87 81 130 113 98 97 99.1
10 126 101 158 129 110 112 107 114 117 95 116.9
11 125 106 104 135 83 177 157 95 101 128 121.1
12 116 134 133 167 142 150 134 153 134 109 137.2
13 53 62 89 76 63 66 69 70 57 49 65.4
14 110 136 75 52 49 85 110 89 68 34 80.8
15 125 112 145 114 61 74 168 98 114 97 110.8
16 87 95 85 111 89 101 85 88 90 164 99.5
17 92 46 74 82 68 94 116 109 81 89 85.1
18 44 74 67 93 89 57 83 53 82 38 68.0
19 95 107 147 154 121 125 146 83 117 139 123.4
20 103 105 139 138 145 145 140 148 137 127 132.7
Table 3.2: Annual losses of the risks
interprets the data as having more uctuations in a single risk than in the mean
values of the risks. This is not surprising because we only have 10 data for each risk
but 20 risks. Thus the uctuations of
Y
i
could also be due to larger uctuations
within the risk.
3.3.2. The B uhlmann Straub model
The main eld of application for credibility theory are collective insurance contracts,
e.g. employees insurance, re insurance for companies, third party liability insurance
for employees, travel insurance for the customers of a travel agent etc.. In reality the
volume of the risk is not the same for each of the contracts. Looking at the credibility
formulae calculated until now we can recognize that the credibility factors should
be larger for higher risk volumes, because we have more data available. Thus we
want to take the volume of the risk into consideration. The risk volume can be the
number of employees, the sum insured, etc..
In order to see how to model dierent risk volumes we rst consider an example.
3. CREDIBILITY THEORY 51
Risk 1 2 3 4 5 6 7
Premium 67.240 100.962 113.336 123.443 160.754 95.578 95.389
Risk 8 9 10 11 12 13 14
Premium 44.570 99.262 116.075 120.043 135.251 67.429 81.976
Risk 15 16 17 18 19 20
Premium 110.313 99.640 86.038 69.885 122.215 131.000
Table 3.3: Credibility premia for the twenty risks
Example 3.2. Assume that we insure taxi drivers. The losses of each driver are
dependent on the driver itself, but also on the policies of the employer and the
region where the company works. The i-th company employs P
i
taxi drivers. There
is a random variable
i
which determines the environment and the policies of the
company. The random variable
ik
determines the risk class of each single driver.
We assume that the vectors (
i1
,
i2
, . . . ,
iP
i
,
i
) are independent and that, given
i
, the parameters
i1
,
i2
, . . . ,
iP
i
are conditionally iid.. The aggregate claims of
dierent companies shall be independent. The aggregate claims of dierent drivers
of company i are conditionally independent given
i
, and the aggregate claims Y
ikj
of a driver are conditionally independent given
ik
and Y
ikj
depends on
ik
only.
Then the expected value of the annual aggregate claims of company i given
i
is
II E
_
P
i
k=1
Y
ikj
i
_
= P
i
II E[Y
i1j
[
i
] .
The conditional variance of company i is then
II E
__
P
i
k=1
Y
ikj
P
i
II E[Y
i1j
[
i
]
_
2
i
_
= P
i
II E
_
(Y
i1j
II E[Y
i1j
[
i
])
2
.
Thus both the conditional mean and the conditional variance are proportional to
characteristics of the company.
The example will motivate the model assumptions. Let P
ij
denote the volume of
the i-th risk in year j. We assume that there exists a parameter
i
for each risk. The
parameters
1
,
2
, . . . ,
n
are assumed to be iid.. Denote by Y
ij
the aggregate claims
of risk i in year j. We assume that the vectors (Y
i1
, Y
i2
, . . . , Y
im
,
i
) are independent.
Within each risk the random variables Y
i1
, Y
i2
, . . . , Y
im
are conditionally independent
given
i
. Moreover, there exist functions m() and s
2
() such that
II E[Y
ij
[
i
= ] = P
ij
m()
52 3. CREDIBILITY THEORY
and
Var[Y
ij
[
i
= ] = P
ij
s
2
() .
As in the B uhlmann model we let = II E[m(
i
)],
2
= II E[s
2
(
i
)] and v
2
=
Var[m(
i
)]. Denote by
P
i
=
m
j=1
P
ij
, P
=
n
i=1
P
i
the risk volume of the i-th risk over all m years and the risk volume over all years
and companies.
First we normalize the annual losses. Let X
ij
= Y
ij
/P
ij
. Note that II E[X
ij
[
i
] =
m(
i
) and Var[X
ij
[
i
] = s
2
(
i
)/P
ij
. We next try to nd the best linear estimator
for m(
i
). Hence we have to minimize
II E
__
a
i0
+
m
j=1
a
ij
X
ij
m(
i
)
_
2
_
.
It is clear that X
kj
does not carry information on m(
i
) if k ,= i. We rst dieren-
tiate with respect to a
i0
.
1
2
d
da
i0
II E
__
a
i0
+
m
j=1
a
ij
X
ij
m(
i
)
_
2
_
= II E
_
a
i0
+
m
j=1
a
ij
X
ij
m(
i
)
_
= a
i0
_
1
m
j=1
a
ij
_
= 0 .
As in the B uhlmann model we get
a
i0
=
_
1
m
j=1
a
ij
_
and therefore
II E
_
a
i0
+
m
j=1
a
ij
X
ij
_
= .
Let us dierentiate with respect to a
ik
,
1
2
d
da
ik
II E
__
a
i0
+
m
j=1
a
ij
X
ij
m(
i
)
_
2
_
= II E
_
X
ik
_
a
i0
+
m
j=1
a
ij
X
ij
m(
i
)
__
= a
i0
+
m
j=1
j=k
a
ij
II E[X
ik
X
ij
] +a
ik
II E
_
X
2
ik
II E[X
ik
m(
i
)] = 0 .
3. CREDIBILITY THEORY 53
Let us compute the terms. For j ,= k
II E[X
ik
X
ij
] = II E[II E[X
ik
X
ij
[
i
]] = II E
_
(m(
i
))
2
= v
2
+
2
.
II E
_
X
2
ik
= II E
_
II E
_
X
2
ik
= II E
_
(m(
i
))
2
+P
1
ik
s
2
(
i
)
= v
2
+
2
+P
1
ik
2
.
II E[X
ik
m(
i
)] = II E[II E[X
ik
m(
i
) [
i
]] = II E
_
(m(
i
))
2
= v
2
+
2
.
Thus
_
1
m
j=1
a
ij
_
2
+
m
j=1
j=k
a
ij
(v
2
+
2
) +a
ik
(v
2
+
2
+P
1
ik
2
) (v
2
+
2
)
= a
ik
P
1
ik
2
_
1
m
j=1
a
ij
_
v
2
= 0 .
The right hand side of
2
P
1
ik
a
ik
=
_
1
m
j=1
a
ij
_
v
2
is independent of k and thus there exists a constant a such that a
ik
= P
ik
a. Then it
follows readily
a
ik
= P
ik
v
2
2
+P
i
v
2
.
Note that the formula is consistent with (3.2). Dene
X
i
=
1
P
i
m
j=1
P
ij
X
ij
the weighted mean of the annual losses of the i-th risk. Then the credibility premium
is
_
1
P
i
v
2
2
+P
i
v
2
_
+
P
i
v
2
2
+P
i
v
2
X
i
= Z
i
X
i
+ (1 Z
i
)
with the credibility factor
Z
i
=
1
1 +
2
P
i
v
2
.
Note that the premium is consistent with (3.3). Denote by
P
i(m+1)
the (estimated)
risk volume of the next year. Then the credibility premium for the next year will be
P
i(m+1)
(Z
i
X
i
+ (1 Z
i
)) .
Note that the credibility factor Z
i
depends on P
i
. This means that in general
dierent risks will get dierent credibility factors.
It remains to nd estimators for the parameters.
54 3. CREDIBILITY THEORY
i) : We want to modify the estimator (3.4) for the B uhlmann Straub model. The
loss X
ij
should be weighted with P
ij
because it originates from a volume of this
size. Thus we get
=
1
P
i=1
m
j=1
P
ij
X
ij
=
1
P
i=1
P
i
X
i
.
It turns out that is unbiased.
ii)
2
: We want to modify the estimator (3.5) for the B uhlmann Straub model.
The conditional variance of X
ij
is P
1
ij
s
2
(
i
). This suggests that we should
weight the term (X
ij
X
i
)
2
with P
ij
. Therefore we try an estimator of the form
c
n
i=1
m
j=1
P
ij
(X
ij
X
i
)
2
.
In order to compute the expected value of this estimator we keep i xed.
II E
_
m
j=1
P
ij
(X
ij
X
i
)
2
_
= II E
_
m
j=1
P
ij
X
2
ij
2
m
j=1
P
ij
X
ij
X
i
+
m
j=1
P
ij
X
2
i
_
= II E
_
m
j=1
P
ij
X
2
ij
_
II E[P
i
X
2
i
]
=
m
j=1
P
ij
_
2
P
ij
+v
2
+
2
_
1
P
i
m
j=1
m
k=1
P
ij
P
ik
II E[X
ij
X
ik
]
= m
2
+P
i
(v
2
+
2
) P
i
(v
2
+
2
)
2
= (m1)
2
. (3.7)
Hence the proposed estimator has mean value
II E
_
c
n
i=1
m
j=1
P
ij
(X
ij
X
i
)
2
_
= cn(m1)
2
.
Thus our unbiased estimator is
2
=
1
n(m1)
n
i=1
m
j=1
P
ij
(X
ij
X
i
)
2
.
iii) v
2
: A closer look at the estimator (3.6) shows hat it contains the terms (Y
ij
)
2
, all with the same weight. For the estimator in the B uhlmann Straub model
we propose an estimator that contains a term of the form
n
i=1
m
j=1
P
ij
(X
ij
)
2
.
3. CREDIBILITY THEORY 55
P
ij
Year
1 2 3 4 5 6 7 8 9 10
1 10 10 12 12 15 15 15 13 9 9
2 5 5 5 5 5 5 5 5 5 5
Company 3 2 2 3 3 3 3 3 3 3 3
number 4 15 20 20 20 21 22 22 22 25 25
5 5 5 5 3 3 3 3 3 3 5
6 10 10 10 10 10 10 10 12 12 12
Table 3.4: Premium volumes
Let us compute its expected value.
II E
_
n
i=1
m
j=1
P
ij
(X
ij
)
2
_
=
n
i=1
m
j=1
P
ij
II E[X
2
ij
] P
II E[
2
]
=
n
i=1
m
j=1
P
ij
_
2
P
ij
+v
2
+
2
_
1
P
i=1
m
j=1
n
k=1
m
l=1
P
ij
P
kl
II E[X
ij
X
kl
]
= nm
2
+P
(v
2
+
2
)
1
P
_
n
i=1
m
j=1
P
2
ij
_
2
P
ij
+v
2
+
2
_
+
n
i=1
m
j=1
m
l=1
l=j
P
ij
P
il
(v
2
+
2
) +
n
i=1
m
j=1
n
k=1
k=i
m
l=1
P
ij
P
kl
2
_
= (nm1)
2
+
_
P
i=1
P
2
i
P
_
v
2
. (3.8)
Let
P
=
1
nm1
n
i=1
P
i
_
1
P
i
P
_
.
Then
v
2
=
1
P
__
1
nm1
n
i=1
m
j=1
P
ij
(X
ij
)
2
_
2
_
is an unbiased estimator of v
2
.
Example 3.3. An insurance company has issued motor insurance policies for
business cars to six companies. Table 3.4 shows the number of cars for each of the
six companies in each of the past ten years. Table 3.5 shows the aggregate claims,
measured in thousands of e, for each of the companies in each of the past ten years.
The rst step is to calculate the mean aggregate claims per car X
ij
= Y
ij
/P
ij
for each
of the companies in each year, see Table 3.6. For the computation of the estimators
56 3. CREDIBILITY THEORY
Y
ij
Year
1 2 3 4 5 6 7 8 9 10
1 74 50 180 43 179 140 95 149 20 81
2 52 111 83 74 87 85 40 71 93 83
Company 3 40 28 60 59 43 74 61 46 72 81
number 4 171 85 100 116 153 44 13 31 110 252
5 49 132 74 37 21 54 27 43 53 102
6 126 148 128 151 165 128 100 65 233 246
Table 3.5: Aggregate annual losses
Year
1 2 3 4 5 6 7 8 9 10
1 7.4 5 15 3.583 11.933 9.333 6.333 11.462 2.222 9
2 10.4 22.2 16.6 14.8 17.4 17 8 14.2 18.6 16.6
3 20 14 20 19.667 14.333 24.667 20.333 15.333 24 27
4 11.4 4.25 5 5.8 7.286 2 0.591 1.409 4.4 10.08
5 9.8 26.4 14.8 12.333 7 18 9 14.333 17.667 20.4
6 12.6 14.8 12.8 15.1 16.5 12.8 10 5.417 19.417 20.5
Table 3.6: Normalised losses
the gures in Table 3.7 are needed. We further get
P
= 554, P
= 7.08585 .
We can now compute the following estimates.
=
1
P
i=1
P
i
X
i
= 9.94765 ,
2
=
1
6
6
i=1
1
9
10
j=1
P
ij
(X
ij
X
i
)
2
= 157.808 ,
v
2
=
1
P
_
1
59
6
i=1
10
j=1
P
ij
(X
ij
)
2
2
_
= 28.7578 .
The number of business cars each company plans to use next year and the credibility
premium is given in Table 3.8. One can clearly see how the credibility factor increases
with P
i
.
3.3.3. The B uhlmann Straub model with missing data
In practice it is not realistic that all policy holders of a certain collective insurance
type started to insure their risks in the same year. Thus some of the P
ij
s may be
3. CREDIBILITY THEORY 57
Company P
i
X
i
j
P
ij
(X
ij
X
i
)
2
j
P
ij
(X
ij
)
2
1 120 8.425 1659.62 1937.84
2 50 15.580 735.78 2321.95
3 28 20.143 494.10 3404.48
4 212 5.071 2310.63 7352.86
5 38 15.579 1289.26 2494.30
6 106 14.057 2032.23 3821.88
Table 3.7: Quantities used in the estimators
Company Cars Credibility Credibility premium Actual premium
next year factor per unit volume
1 9 0.95627 8.4916 76.424
2 6 0.90110 15.0230 90.138
3 3 0.83613 18.4722 55.417
4 24 0.97477 5.1938 124.651
5 3 0.87382 14.8684 44.605
6 12 0.95078 13.8544 166.252
Table 3.8: Premium next year
0. The question arises whether this has an inuence on the model. The answer is
that its inuence is only small. Denote by m
i
the number of years with P
ij
,= 0.
For the credibility estimator it is clear that a
ij
= 0 if P
ij
= 0 because then also
Y
ij
= 0. For convenience we dene X
ij
= 0 if P
ij
= 0. The computation of a
ij
does
not change if some of the P
ij
s are 0. Moreover, it is easy to see that remains
unbiased.
Next let us consider the estimator of
2
. A closer look at (3.7) shows that
II E
_
m
j=1
P
ij
(X
ij
X
i
)
2
_
= (m
i
1)
2
.
Hence the estimator of
2
must be changed to
2
=
1
n
n
i=1
1
m
i
1
m
j=1
P
ij
(X
ij
X
i
)
2
.
Consider the estimator of v. The computation (3.8) changes to
II E
_
n
i=1
m
j=1
P
ij
(X
ij
)
2
_
= (m
1)
2
+
_
P
i=1
P
2
i
P
_
v
2
58 3. CREDIBILITY THEORY
where
m
=
n
i=1
m
i
.
Thus one has to change
P
=
1
m
1
n
i=1
P
i
_
1
P
i
P
_
and
v
2
=
1
P
__
1
m
1
n
i=1
m
j=1
P
ij
(X
ij
)
2
_
2
_
.
3.4. General Bayes Methods
We can generalise the approach above by summarise the observations in a vector X
i
.
In the models considered before we would have X
i
= (X
i,1
, X
i,2
, . . . , X
i,m
i
)
. We
then want to estimate the vector m
i
IR
s
. In the models above we had m
i
= m(
i
).
We assume that the vectors (X
i
, m
i
)
and X
i
= (X
i1
, X
2
i1
, X
i2
, X
2
i2
, . . . , X
i,m
i
, X
2
i,m
i
)
.
The goal is now to estimate m
i
.
The estimator M minimising II E[|M m
i
|
2
] is then M = II E[m
i
[ X
i
]. The
problem with the estimator is again, that we need the joint distribution of the
quantities. We therefore restrict again to linear estimators, i.e. estimators of the
form M = g + GX, where we for the moment omit the index. We denote the
entries of g by (g
i0
). The quantity to minimise is then
(M) :=
i
II E
__
g
i0
+
m
j=1
g
ij
X
j
m
i
_
2
_
.
In order to minimise (M) we rst take the derivative with respect to g
i0
and equate
it to zero
d(M)
dg
i0
= 2II E
_
g
i0
+
m
=1
g
i
X
m
i
_
= 0 .
The derivative with respect to g
ij
yields
d(M)
dg
ij
= 2II E
_
X
j
_
g
i0
+
m
=1
g
i
X
m
i
__
= 0 .
3. CREDIBILITY THEORY 59
In matrix form we can express the equations as
II E[g +GX m] = 0,
II E[(g +GX m)X
] = 0.
The solution we look for is therefore
g = II E[m] GII E[X] , (3.9a)
G = Cov[m, X
] Var[X]
1
. (3.9b)
Here we assume that Var[X] is invertible. Note that Var[X] is symmetric. Because
for any vector a we have a
Var[X]a = Var[a
Var[X]a = 0,
or equivalently, a
]
becomes invertible.
If we now again consider n collective contracts (X
i
, m
i
) the linear Bayes es-
timator M
i
for m
i
can be written as
M
i
= II E[m
i
] + Cov[m
i
, X
i
] Var[X
i
]
1
(X
i
II E[X
i
]) .
Thus we start with expected value II E[m
i
] and correct it according to the deviation
of X
i
from the expected value II E[X
i
]. The correction is larger if the covariance
between m
i
and X
i
is larger, smaller if the variance of X
i
is larger.
Example 3.4. Suppose there is an unobserved variable
i
that determines the
risk. Suppose II E[X
i
[
i
] = Y
i
b(
i
) and Var[X
i
[
i
] = P
i
V
i
(
i
) for some known
Y
i
IR
mq
, P
i
IR
mm
and some functions b() IR
q
, V
i
() IR
mm
of the
unobservable random variable
i
. We assume that P
i
is invertible. The B uhlmann
and the B uhlmann-Straub models are special cases.
We here are interested to estimate b(
i
). Introduce the following quantities
= II E[b(
i
)], = Var[b(
i
)] and
i
= P
i
II E[V
i
(
i
)]. The moments required are
, II E[X
i
] = Y
i
,
Cov[b(
i
), X
i
] = Cov[II E[b(
i
) [
i
], II E[X
i
[
i
]] + II E[Cov[b(
i
), X
i
[
i
]]
= Cov[b(
i
), b(
i
)
i
] = Cov[b(
i
), b(
i
)
]Y
i
= Y
i
,
and
Var[X
i
] = Var[II E[X
i
[
i
]] + II E[Var[X
i
[
i
]] = Y
i
Y
i
+
i
.
60 3. CREDIBILITY THEORY
We assume that II E[V (
i
)] and therefore
i
is invertible. This yields the solution
b
i
= Y
i
(Y
i
Y
i
+
i
)
1
X
i
+ (I Y
i
(Y
i
Y
i
+
i
)
1
Y
i
) . (3.10)
Note that is symmetric. If would not be invertible could we write some co-
ordinates of b(
i
) as a linear function of the others. Thus we assume that is
invertible. A well know formula gives
(Y
i
Y
i
+
i
)
1
=
1
i
1
i
Y
i
(
1
+Y
i
1
i
Y
i
)
1
Y
i
1
i
.
Dene the matrix
Z
i
= Y
i
1
i
Y
i
(
1
+Y
i
1
i
Y
i
)
1
1
= Y
i
1
i
Y
i
(Y
i
1
i
Y
i
+I)
1
.
Then we nd
Z
i
= (Y
i
1
i
Y
i
+I I)(Y
i
1
i
Y
i
+I)
1
= I (Y
i
1
i
Y
i
+I)
1
.
We also nd
Y
i
(Y
i
Y
i
+
i
)
1
Y
i
= Y
i
1
i
Y
i
Y
i
1
i
Y
i
(
1
+Y
i
1
i
Y
i
)
1
Y
i
1
i
Y
i
= (I Z
i
)Y
i
1
i
Y
i
= (Y
i
1
i
Y
i
+I)
1
Y
i
1
i
Y
i
= Z
i
.
It follows that we can write the LB estimator as
b
i
= (I Z
i
)(Y
i
1
i
X
i
+) . (3.11)
The matrix Z
i
is called the credibility matrix.
Suppose that in addition Y
i
1
i
Y
i
is invertible. This is equivalent to that Y
i
has full rank q and m q. Then
(Y
i
1
i
Y
i
+I)(Y
i
1
i
Y
i
)
1
Y
i
1
i
= (+ (Y
i
1
i
Y
i
)
1
)Y
i
1
i
.
Multiplying by (Y
i
1
i
Y
i
+I)
1
from the left gives
(Y
i
1
i
Y
i
)
1
Y
i
1
i
= (Y
i
1
i
Y
i
+I)
1
(+ (Y
i
1
i
Y
i
)
1
)Y
i
1
i
.
Rearranging the term yields
(I Z
i
)Y
i
1
i
= (Y
i
1
i
Y
i
+I)
1
Y
i
1
i
= (I (Y
i
1
i
Y
i
+I)
1
)(Y
i
1
i
Y
i
)
1
Y
i
1
i
= Z
i
(Y
i
1
i
Y
i
)
1
Y
i
1
i
.
3. CREDIBILITY THEORY 61
If we dene
b
i
= (Y
i
1
i
Y
i
)
1
Y
i
1
i
X
i
we can express the estimator in credibility weighted form
b
i
= Z
i
b
i
+ (I Z
i
) .
we have (M) =
|mm
/
|
2
+|m
/
M|
2
which clearly is minimised by choosing M = m
/
.
If L
t
L is a closed linear subspace then iterated projections give m
/
=
pro(m
/
[ L
t
). More generally, if 0 = L
0
L
1
L
n
= L
2
is a nested family
of closed linear subspaces of L
2
. Then we have m = m
/n
and
m
/
k
=
k
j=1
(m
/
j
m
/
j1
) .
For exampe if L
1
is the space of constants and L
2
is the space of linear estimators
we nd m
/
1
= II E[m
i
] and
m
/
2
m
/
1
= Cov[m
i
, X
I
] Var[X
i
]
1
(X
i
II E[X
i
]) .
Bibliographical Remarks
Credibility theory originated in North America in the early part of the twentieth
century. These early ideas are now known as American credibility. References can
be found in the survey paper by Norberg [63]. The Bayesian approach to credibility
is generally attributed to Bailey [14] and [15]. The empirical Bayes approach is due
to B uhlmann [20] and B uhlmann and Straub [21].
62 4. THE CRAM
ER-LUNDBERG MODEL
4. The Cramer-Lundberg Model
4.1. Denition of the Cramer-Lundberg Process
We have seen that the compound Poisson model has nice properties. For instance it
can be derived as a limit of individual models. This was the reason for Filip Lundberg
to postulate a continuous time risk model where the aggregate claims in any interval
have a compound Poisson distribution. Moreover, the premium income should be
modelled. In a portfolio of insurance contracts the premium payments will be spread
all over the year. Thus he assumed that that the premium income is continuous over
time and that the premium income in any time interval is proportional to the interval
length. This leads to the following model for the surplus of an insurance portfolio
C
t
= u +ct
Nt
i=1
Y
i
.
u is the initial capital, c is the premium rate. The number of claims in (0, t] is a
Poisson process N
t
with rate . The claim sizes Y
i
are a sequence of iid. positive
random variables independent of N
t
. This model is called Cramer-Lundberg
process or classical risk process.
We denote the distribution function of the claims by G, its moments by
n
=
II E[Y
n
1
] and its moment generating function by M
Y
(r) = II E[exprY
1
]. Let =
1
.
We assume that < . Otherwise no insurance company would insure such a risk.
Note that G(x) = 0 for x < 0. We will see later that it is no restriction to assume
that G(0) = 0.
For an insurance company it is important that C
t
stays above a certain level.
This level is given by legal restrictions. By adjusting the initial capital it is no loss
of generality to assume this level to be 0. We dene the ruin time
= inft > 0 : C
t
< 0 , (inf = ) .
We will mostly be interested in the probability of ruin in a time interval (0, t]
(u, t) = II P[ t [ C
0
= u] = II P[ inf
0<st
C
s
< 0 [ C
0
= u]
and the probability of ultimate ruin
(u) = lim
t
(u, t) = II P[inf
t>0
C
t
< 0 [ C
0
= u] .
4. THE CRAM
ER-LUNDBERG MODEL 63
It is easy to see that (u, t) is decreasing in u and increasing in t.
We denote the claim times by T
1
, T
2
, . . . and by convention T
0
= 0. Let X
i
=
c(T
i
T
i1
) Y
i
. If we only consider the process at the claim times we can see that
C
Tn
= u +
n
i=1
X
i
is a random walk. Note that (u) = II P[inf
nIIN
C
Tn
< 0]. From the theory of random
walks we can see that ruin occurs a.s. i II E[X
i
] 0, compare with Lemma E.1 or
[40, p.396]. Hence we will assume in the sequel that
II E[X
i
] > 0 c
1
i=1
Y
i
_
= t.
The condition can be interpreted that the mean income is strictly larger than the
mean outow. Therefore the condition is also called the net prot condition.
If the net prot condition is fullled then C
Tn
tends to innity as n . Hence
infC
t
u : t > 0 = infC
Tn
u : n 1
is a.s. nite. So we can conclude that
lim
u
(u) = 0 .
4.2. A Note on the Model and Reality
In reality the mean number of claims in an interval will not be the same all the
time. There will be a claim rate (t) which may be periodic in time. Moreover,
the number of individual contracts in the portfolio may vary with time. Let a(t) be
the volume of the portfolio at time t. Then the claim number process N
t
is an
inhomogeneous Poisson process with rate a(t)(t). Let
(t) =
_
t
0
a(s)(s) ds
and
1
(t) be its inverse function. Then
N
t
= N
1
(t)
is a Poisson process with rate
1. Let now the premium rate vary with t such that c
t
= ca(t)(t) for some constant
c. This assumption is natural for changes in the risk volume. It is articial for the
64 4. THE CRAM
ER-LUNDBERG MODEL
changes in the intensity. For instance we assume that the company gets more new
customers at times with a higher intensity. This eect may arise because customers
withdraw from their old insurance contracts and write new contracts with another
company after claims occurred because they were not satised by the handling of
claims by their old companies.
The premium income in the interval (0, t] is c(t). Let
C
t
= C
1
(t)
. Then
C
t
= u +c(
1
(t))
N
1
(t)
i=1
Y
i
= u +ct
Nt
i=1
Y
i
is a Cramer-Lundberg process. Thus we should not consider time to be the real
time but operational time.
The event of ruin does almost never occur in practice. If an insurance company
observes that their surplus is decreasing they will immediately increase their premia.
On the other hand an insurance company is built up on dierent portfolios. Ruin in
one portfolio does not mean bankruptcy. Therefore ruin is only a technical term. The
probability of ruin is used for decision taking, for instance the premium calculation
or the computation of reinsurance retention levels. For an actuary it is important
to be able to take a good decision in reasonable time. Therefore it is ne to simplify
the model in order to be able to check whether a decision has the desired eect or
not.
The surplus will also be a technical term in practice. If the business is going well
then the share holders will decide to get a higher dividend. To model this we would
have to assume a premium rate dependent on the surplus. But then it would be
hard to obtain any useful results. For some references see for instance [42] and [11].
In Section 1 about risk models we saw that a negative binomial distribution
for the number of claims in a certain time interval would be preferable. We will
later consider such models. But actuaries use the Cramer-Lundberg model for their
calculations, even though there is a certain lack of reality. The reason is that this is
more or less the only model which they can handle.
4.3. A Dierential Equation for the Ruin Probability
We rst prove that C
t
is a strong Markov process.
Lemma 4.1. Let C
t
be a Cramer-Lundberg process and T be a nite stopping
time. Then the stochastic process C
T+t
C
T
: t 0 is a Cramer-Lundberg process
with initial capital 0 and independent of T
T
.
4. THE CRAM
ER-LUNDBERG MODEL 65
Proof. We can write C
T+t
C
T
as
ct
N
T+t
i=N
T
+1
Y
i
.
Because the claim amounts are iid. and independent of N
t
we only have to prove
that N
T+t
N
T
is a Poisson process independent of T
T
. Because N
t
is a renewal
process it is enough to show that T
N
T
+1
T is Exp() distributed and independent
of T
T
. Condition on T
N
T
and T. Then
II P[T
N
T
+1
T > x [ T
N
T
, T]
= II P[T
N
T
+1
T
N
T
> x +T T
N
T
[ T
N
T
, T, T
N
T
+1
T
N
T
> T T
N
T
] = e
x
by the lack of memory property of the exponential distribution. The assertion
follows because T
N
T
+1
depends on T
T
via T
N
T
and T only. The latter, even though
intuitively clear, follows readily noting that T
T
is generated by sets of the form
N
T
= n, A
n
where n IIN and A
n
(Y
1
, . . . , Y
n
, T
1
, . . . , T
n
).
Let h be small. If ruin does not occur in the interval (0, T
1
h] then a new
Cramer-Lundberg process starts at time T
1
h with new initial capital C
T
1
h
. Let
(u) = 1 (u) denote the probability that ruin does not occur. Using that the
interarrival times are exponentially distributed we have the density e
t
of the
distribution of T
1
and II P[T
1
> h] = e
h
. We obtain noting that (x) = 0 for x < 0
(u) = e
h
(u +ch) +
_
h
0
_
u+ct
0
(u +ct y) dG(y)e
t
dt .
Letting h tending to 0 shows that (u) is right continuous. Rearranging the terms
and dividing by h yields
c
(u +ch) (u)
ch
=
1 e
h
h
(u +ch)
1
h
_
h
0
_
u+ct
0
(u +ct y) dG(y)e
t
dt .
Letting h tend to 0 shows that (u) is dierentiable from the right and
c
t
(u) =
_
(u)
_
u
0
(u y) dG(y)
_
. (4.1)
Replacing u by u ch gives
(u ch) = e
h
(u) +
_
h
0
_
uc(ht)
0
(u (c h)t y) dG(y)e
t
dt .
66 4. THE CRAM
ER-LUNDBERG MODEL
We conclude that (u) is continuous. Rearranging the terms, dividing by h and
letting h 0 yields the equation
c
t
(u) =
_
(u)
_
u
0
(u y) dG(y)
_
.
We see that (u) is dierentiable at all points u where G(u) is continuous. If G(u)
has a jump then the derivatives from the left and from the right do not coincide.
But we have shown that (u) is absolutely continuous. If we now let
t
(u) be the
derivative from the right we have that
t
(u) is a density of (u).
The diculty with equation (4.1) is that it contains both the derivative of (u)
and an integral. Let us try to get rid of the derivative. We nd
c
((u) (0)) =
1
_
u
0
c
t
(x) dx =
_
u
0
(x) dx
_
u
0
_
x
0
(x y) dG(y) dx
=
_
u
0
(x) dx
_
u
0
_
u
y
(x y) dx dG(y)
=
_
u
0
(x) dx
_
u
0
_
uy
0
(x) dx dG(y)
=
_
u
0
(x) dx
_
u
0
_
ux
0
dG(y) (x) dx
=
_
u
0
(x)(1 G(u x)) dx =
_
u
0
(u x)(1 G(x)) dx
Note that
_
0
(1G(x)) dx = . Letting u we can by the bounded convergence
theorem interchange limit and integral and get
c(1 (0)) =
_
0
(1 G(x)) dx =
where we used that (u) 1. It follows that
(0) = 1
c
, (0) =
c
.
Replacing (u) by 1 (u) we obtain
c(u) =
_
u
0
(1 (u x))(1 G(x)) dx
=
__
u
(1 G(x)) dx +
_
u
0
(u x)(1 G(x)) dx
_
. (4.2)
In Section 4.8 we shall obtain a natural interpretation of (4.2).
4. THE CRAM
ER-LUNDBERG MODEL 67
Example 4.2. Let the claims be Exp() distributed. Then the equation (4.1) can
be written as
c
t
(u) =
_
(u) e
u
_
u
0
(y)e
y
dy
_
.
Dierentiating yields
c
tt
(u) =
_
t
(u) +e
u
_
u
0
(y)e
y
dy (u)
_
=
t
(u) c
t
(u) .
The solution to this dierential equation is
(u) = A +Be
(/c)u
.
Because (u) 1 as u we get A = 1. Because (0) = 1 /(c) the solution
is
(u) = 1
c
e
(/c)u
or
(u) =
c
e
(/c)u
.
68 4. THE CRAM
ER-LUNDBERG MODEL
R
r
r
1.0 0.5 0.5 1.0 1.5
0.2
0.4
0.6
0.8
Figure 4.1: The function (r) and the adjustment coecient R
It would be nice to have a martingale that only depends on the surplus and not
explicitly on time. Let us therefore consider the equation (r) = 0. This equation
has obviously the solution r = 0. We dierentiate (r).
t
(r) = M
t
Y
(r) c ,
tt
(r) = M
tt
Y
(r) = II E[Y
2
e
rY
] > 0 .
The function (r) is strictly convex. For r = 0
t
(0) = M
t
Y
(0) c = c < 0
by the net prot condition. There might be at most one additional solution R to
the equation (R) = 0 and R > 0. If this solution exists we call it the adjustment
coecient or the Lundberg exponent. The Lundberg exponent will play an
important role in the estimation of the ruin probabilities.
Example 4.2 (continued). For Exp() distributed claims we have to solve
_
r
1
_
cr =
r
r
cr = 0 .
For r ,= 0 we nd R = /c. Note that
(u) =
c
e
Ru
.
4. THE CRAM
ER-LUNDBERG MODEL 69
In general the adjustment coecient is dicult to calculate. So we try to nd
some bounds for the adjustment coecient. Note that the second moment
2
exists
if the Lundberg exponent exists. Consider the function (r) for r > 0.
tt
(r) = II E[Y
2
e
rY
] > II E[Y
2
] =
2
,
t
(r) =
t
(0) +
_
r
0
tt
(s) ds > (c ) +
2
r ,
(r) = (0) +
_
r
0
t
(s) ds >
2
r
2
2
(c )r .
The last inequality yields for r = R
0 = (R) > R
_
2
R
2
(c )
_
from which an upper bound of R
R <
2(c )
2
follows.
We are not able to nd a lower bound in general. But in the case of bounded
claims we get a lower bound for R. Assume that Y
1
M a.s..
Let us rst consider the function
f(x) =
x
M
_
e
RM
1
_
_
e
Rx
1
_
.
Its second derivative is
f
tt
(x) = R
2
e
Rx
< 0 .
f(x) is concave with f(0) = f(M) = 0. Thus f(x) > 0 for 0 < x < M. The
function h(x) = xe
x
e
x
+ 1 has a minimum in 0. Thus h(x) > h(0) = 0 for x ,= 0,
in particular
1
RM
_
e
RM
1
_
< e
RM
.
We calculate
M
Y
(R) 1 =
_
M
0
_
e
Rx
1
_
dG(x) <
_
M
0
x
M
_
e
RM
1
_
dG(x) =
M
_
e
RM
1
_
.
From the equation determining R we get
0 = (M
Y
(R) 1) cR <
M
_
e
RM
1
_
cR < Re
RM
cR.
It follows readily
R >
1
M
log
c
.
70 4. THE CRAM
ER-LUNDBERG MODEL
4.5. Lundbergs Inequality
We will now connect the adjustment coecient and ruin probabilities.
Theorem 4.4. Assume that the adjustment coecient R exists. Then
(u) < e
Ru
.
Proof. Assume that the theorem does not hold. Let
u
0
= infu 0 : (u) e
Ru
.
Because (u) is continuous we get
(u
0
) = e
Ru
0
.
Because (0) < 1 we conclude that u
0
> 0. Consider the equation (4.2) for u = u
0
.
c(u
0
) =
_
_
u
0
(1 G(x)) dx +
_
u
0
0
(u
0
x)(1 G(x)) dx
_
<
_
_
u
0
(1 G(x)) dx +
_
u
0
0
e
R(u
0
x)
(1 G(x)) dx
_
_
0
e
R(u
0
x)
(1 G(x)) dx = e
Ru
0
_
0
_
x
e
Rx
dG(y) dx
= e
Ru
0
_
0
_
y
0
e
Rx
dx dG(y) = e
Ru
0
_
0
1
R
(e
Ry
1) dG(y)
=
R
e
Ru
0
(M
Y
(R) 1) = ce
Ru
0
which is a contradiction. This proves the theorem.
We now give an alternative proof of the theorem. We will use the positive
martingale (4.4) for r = R. By the stopping theorem
e
Ru
= e
RC
0
= II E[e
RCt
] II E[e
RCt
; t] = II E[e
RC
; t] .
Letting t tend to innity yields by monotone convergence
e
Ru
II E[e
RC
; < ] > II P[ < ] = (u) (4.5)
because C
< 0. We have got an upper bound for the ruin probability. We would
like to know whether R is the best possible exponent in an exponential upper bound.
This question will be answered in the next section.
4. THE CRAM
ER-LUNDBERG MODEL 71
Example 4.5. Let G(x) = 1 pe
x
(1 p)e
x
where 0 < < and 0 < p
( )
1
. Note that the mean value is
p
+
1p
and thus
c >
p
+
(1 p)
.
For r <
M
Y
(r) =
_
0
e
rx
_
pe
x
+(1 p)e
x
_
dx =
p
r
+
(1 p)
r
.
Thus we have to solve
_
p
r
+
(1 p)
r
1
_
cr =
_
pr
r
+
(1 p)r
r
_
cr = 0 .
We nd the obvious solution r = 0. If r ,= 0 then
p( r) +(1 p)( r) = c( r)( r)
or equivalently
cr
2
(( +)c )r +c ((1 p) +p) = 0 . (4.6)
The solution is
r =
1
2
_
+
c
_
_
+
c
_
2
4
_
c
p
c
(1 p)
_
_
=
1
2
_
+
c
_
_
c
_
2
+ 4p
c
( )
_
.
Now we got three solutions. But there should only be two. Note that
c
p
c
(1 p) =
c
_
c
_
p
+
1 p
__
> 0
and thus both solutions are positive. The larger of the solutions can be written as
+
1
2
_
c
+
_
_
c
_
2
+ 4p
c
( )
_
and is thus larger than . But the moment generating function does not exist for
r . Thus
R =
1
2
_
+
c
_
_
c
_
2
+ 4p
c
( )
_
.
From Lundbergs inequality it follows that
(u) < e
Ru
.
72 4. THE CRAM
ER-LUNDBERG MODEL
4.6. The Cramer-Lundberg Approximation
Consider the equation (4.2). This equation looks almost like a renewal equation,
but
_
0
c
(1 G(x)) dx =
c
< 1
is not a probability distribution. Can we manipulate (4.2) to get a renewal equation?
We try for some measurable function h and
h
(u)h(u) = h(u)
_
u
c
(1G(x)) dx+
_
u
0
(ux)h(ux)
h(x)
c
(1G(x)) dx (4.7)
where h(u) = h(ux)
c
(1 G(x)) dx =
c
_
0
_
x
e
rx
dG(y) dx =
(M
Y
(r) 1)
cr
.
The only solution to the latter equation is r = R. Let us assume that R exists.
Moreover, we need that
_
0
xe
Rx
c
(1 G(x)) dx < (4.8)
which is equivalent to M
t
Y
(R) < , see below. The equation is now
(u)e
Ru
= e
Ru
_
u
c
(1 G(x)) dx +
_
u
0
(u x)e
R(ux)
e
Rx
c
(1 G(x)) dx .
It can be shown that
e
Ru
_
u
c
(1 G(x)) dx
is directly Riemann integrable. Thus we get from the renewal theorem
Theorem 4.6. Assume that the Lundberg exponent exists and that (4.8) is fullled.
Then
lim
u
(u)e
Ru
=
c
M
t
Y
(R) c
.
Proof. It only remains to compute the limit.
_
0
e
Ru
_
u
c
(1 G(x)) dx du =
c
_
0
_
x
0
e
Ru
du(1 G(x)) dx
=
cR
_
0
(e
Rx
1)(1 G(x)) dx =
1
R
cR
=
1
cR
(c ) .
4. THE CRAM
ER-LUNDBERG MODEL 73
The mean value of the distribution is
_
0
xe
Rx
c
(1 G(x)) dx =
c
_
0
_
x
xe
Rx
dG(y) dx
=
c
_
0
_
y
0
xe
Rx
dx dG(y) =
cR
2
_
0
(Rye
Ry
e
Ry
+ 1) dG(y)
=
(RM
t
Y
(R) M
Y
(R) + 1)
cR
2
=
RM
t
Y
(R) cR
cR
2
=
M
t
Y
(R) c
cR
.
The limit value follows readily.
Remark. The theorem shows that it is not possible to obtain an exponential
upper bound for the ruin probability with an exponent strictly larger than R.
The theorem can be written in the following form
(u)
c
M
t
Y
(R) c
e
Ru
.
Thus for large u we get an approximation to (u). This approximation is called the
Cramer-Lundberg approximation.
Example 4.2 (continued). From M
Y
(r) we get that
M
t
Y
(r) =
( r)
2
and thus
lim
u
(u)e
Ru
=
c
(R)
2
c
=
c
c
)
2
c
=
c
c
c
=
c
.
Hence the Cramer-Lundberg approximation
(u)
c
e
(/c)u
becomes exact in this case.
Example 4.7. Let c = = 1 and G(x) = 1
1
3
(e
x
+ e
2x
+ e
3x
). The mean
value of claim sizes is = 0.611111, i.e. the net prot condition is fullled. One can
show that
(u) = 0.550790 e
0.485131u
+ 0.0436979 e
1.72235u
+ 0.0166231 e
2.79252u
.
From Theorem 4.6 it follows that app(u) = 0.550790 e
0.485131u
is the Cramer-
Lundberg approximation to (u). Table 4.1 below shows the ruin function (u), its
Cramer-Lundberg approximation app(u) and the relative error (app(u)(u))/(u)
multiplied by 100 (Er). Note that the the relative error is below 1% for u
1.71358 = 2.8.
74 4. THE CRAM
ER-LUNDBERG MODEL
u 0 0.25 0.5 0.75 1
(u) 0.6111 0.5246 0.4547 0.3969 0.3479
app(u) 0.5508 0.4879 0.4322 0.3828 0.3391
Er -9.87 -6.99 -4.97 -3.54 -2.54
u 1.25 1.5 1.75 2 2.25
(u) 0.3059 0.2696 0.2379 0.2102 0.1858
app(u) 0.3003 0.2660 0.2357 0.2087 0.1849
Er -1.82 -1.32 -0.95 -0.69 -0.50
Table 4.1: Cramer-Lundberg approximation to ruin probabilities
4.7. Reinsurance and Ruin
4.7.1. Proportional Reinsurance
Recall that for proportional insurance the insurer covers Y
I
i
= Y
i
of each claim,
the reinsurer covers Y
R
i
= (1 )Y
i
. Denote by c
I
the insurers premium rate. The
insurers adjustment coecient is obtained from the equation
(M
Y
(r) 1) c
I
r = 0 .
The new moment generating function is
M
Y
(r) = II E[e
rY
i
] = M
Y
(r) .
Assume that both insurer and reinsurer use an expected value premium principle
with the same safety loading. Then c
I
= c and we have to solve
(M
Y
(r) 1) cr = 0 .
This is almost the original equation, hence R = R
I
where R
I
is the adjustment
coecient under reinsurance. The new adjustment coecient is larger, hence the
risk has become smaller.
Example 4.8. Let Y
i
Exp() and let the gross premium rate be xed (1 +
)II E[Y ] = (1 + )
. We
assume that , otherwise the insurer could choose = 0 and he would make a
prot without any risk. How shall we choose in order to maximize the insurers
adjustment coecient? The case = had been considered before. We only have
4. THE CRAM
ER-LUNDBERG MODEL 75
to consider the case > . Because the net prot condition must be fullled there
is a lower bound for .
c
I
= (1 +)
(1 +)(1 )
= ((1 +) ( ))
>
.
Thus
>
= 1
. (4.9)
The moment generating function of Y is
M
Y
(r) = M
Y
(r) =
r
=
r
.
The claims after reinsurance are Exp(/) distributed and the adjustment coecient
is
R
I
() =
((1 +) ( ))
=
_
1
1
(1 +) ( )
_
.
Setting the derivative to 0 yields
2
=
((1 +) ( ))
2
1 +
= (1 +)
2
2( ) +
( )
2
1 +
.
The solution to this equation is
=
1
_
_
( )
2
1 +
( )
2
_
=
_
1
_
_
1
_
1
1 +
_
=
_
1
_
_
1 +
_
1
1 +
_
where the last equality follows from (4.9). This solution is strictly larger than 1 if
>
2
+ 2. Thus the solution is
max
= min
_
_
1
_
_
1 +
_
1
1 +
_
, 1
_
.
By plugging in the values 1
__
M
0
e
rx
dG(x) + e
rM
(1 G(M)) 1
_
c
I
r = 0 .
76 4. THE CRAM
ER-LUNDBERG MODEL
There is no possibility to nd the solution from the problem without reinsurance.
We have to solve the equation for every M separately. But note that R
I
exists
in any case. Especially for heavy tailed distributions this shows that the risk has
become much smaller. By the Cramer-Lundberg approximation the ruin probability
decreases exponentially as the initial capital increases.
We will now show that for the insurer the excess of loss reinsurance is optimal.
We assume that both insurer and reinsurer use an expected value principle.
Proposition 4.9. Let all premia be computed via the expected value principle.
Under all reinsurance forms acting on individual claims with premium rates c
I
and
c
R
xed the excess of loss reinsurance maximizes the insurers adjustment coecient.
Proof. Let h(x) be an increasing function with 0 h(x) x for x 0. We
assume that the insurer pays Y
I
i
= h(Y
i
). Let h
(Y
i
)] = II E[h(Y
i
)] =
_
0
h(y) dG(y) .
Because e
z
1 +z we obtain
e
r(h(y)h
(y))
1 +r(h(y) h
(y))
and
e
rh(y)
e
rh
(y)
(1 +r(h(y) h
(y))) .
Thus
M
h(Y )
(r) =
_
0
e
rh(y)
dG(y)
_
0
e
rh
(y)
(1 +r(h(y) h
(y))) dG(y)
= M
h
(Y )
(r) +r
_
0
(h(y) h
(y)) e
rh
(y)
dG(y) .
For y U we have h(y) y = h
(y))e
rh
(y)
dG(y)
=
_
U
0
(h(y) h
(y))e
rh
(y)
dG(y) +
_
U
(h(y) h
(y))e
rh
(y)
dG(y)
_
U
0
(h(y) h
(y))e
rU
dG(y) +
_
U
(h(y) h
(y))e
rU
dG(y)
= e
rU
_
0
(h(y) h
(y)) dG(y)
= e
rU
(II E[h(Y )] II E[h
(Y )]) = 0 .
4. THE CRAM
ER-LUNDBERG MODEL 77
It follows that M
h(Y )
(r) M
h
(Y )
(r) for r > 0. Since
0 = (R
I
) = (M
h(Y )
(R
I
) 1) c
I
R
I
(M
h
(Y )
(R
I
) 1) c
I
R
I
=
(R
I
)
and the assertion follows from the convexity of
(r).
We consider now the portfolio of the reinsurer. What is the claim number process
of the claims the reinsurer is involved? Because the claim amounts are independent
of the claim arrival process we delete independently points from the Poisson process
with probability G(M) and do not delete them with probability 1 G(M). By
Proposition C.3 this process is a Poisson process with rate (1 G(M)). Because
the claim sizes are iid. and independent of the claim arrival process the surplus of
the reinsurer is a Cramer-Lundberg process with intensity (1 G(M)) and claim
size distribution
G(x) = II P[Y
i
M x [ Y
i
> M] =
G(M +x) G(M)
1 G(M)
.
4.8. The Severity of Ruin and the Distribution of infC
t
: t 0
For an insurance company ruin is not so dramatic if C
x
(u) = II P[ < , C
< x] .
We proceed as in Section 4.3. For h small
x
(u) = e
h
x
(u+ch)+
_
h
0
_
_
u+ct
0
x
(u+cty) dG(y)+(1G(u+ct+x))
_
e
t
dt .
It follows that
x
(u) is right continuous. We obtain the integro-dierential equation
c
t
x
(u) =
_
x
(u)
_
u
0
x
(u y) dG(y) (1 G(u +x))
_
.
Replacing u by u ch gives the derivative from the left. Integration of the integro-
dierential equation yields
c
(
x
(u)
x
(0)) =
_
u
0
x
(u y)(1 G(y)) dy
_
u
0
(1 G(y +x)) dy
=
_
u
0
x
(u y)(1 G(y)) dy
_
x+u
x
(1 G(y)) dy .
78 4. THE CRAM
ER-LUNDBERG MODEL
Because
x
(u) (u) we can see that
x
(u) 0 as u . By the bounded
convergence theorem we obtain
x
(0) =
_
x
(1 G(y)) dy
and
II P[C
< x [ < , C
0
= 0] =
c
_
x
(1 G(y)) dy
c
=
1
_
x
(1 G(y)) dy .
(4.10)
The random variable C
1
0.
Ruin occurs at
1
.
Thus we get
(u) =
_
1
c
_
0 +
c
_
u
0
(u y)(1 G(y)) dy +
c
_
u
1 (1 G(y)) dy .
This is equation (4.2). Thus we have now a natural interpretation of (4.2).
Let
0
= 0 and
i
= inft >
i1
: C
t
< C
i1
, called the ladder times,
and dene L
i
= C
i1
C
i
, called the ladder heights. Note that L
i
only is
dened if
i
< . By Lemma 4.1 we nd II P[
i
< [
i1
< ] = /c. Let
K = supi IIN :
i
< be the number of ladder epochs. We have just seen that
K NB(1, 1 /c) and that, given K, the random variables (L
i
: i K) are
iid. and absolutely continuous with density (1G(x))/. We only have to condition
on K because L
i
is not dened for i > K. If we assume that all (L
i
: i 1) have the
same distribution, then we can drop the conditioning on K and (L
i
) is independent
of K. Then
infC
t
: t 0 = u
K
i=1
L
i
4. THE CRAM
ER-LUNDBERG MODEL 79
and
II P[ < ] = II P[infC
t
: t 0 < 0] = II P
_
K
i=1
L
i
> u
_
.
Denote by
B(x) =
1
_
x
0
(1 G(y)) dy
the distribution function of L
i
. We can use Panjer recursion to approximate (u)
by using an appropriate discretization.
A formula that is useful for theoretical considerations, but too messy to use for
the computation of (u) is the Pollaczek-Khintchine formula.
(u) = II P
_
K
i=1
L
i
> u
_
=
n=1
II P
_
n
i=1
L
i
> u
_
II P[K = n]
=
_
1
c
_
n=1
_
c
_
n
(1 B
n
(u)) . (4.11)
4.9. The Laplace Transform of
Denition 4.10. Let f be a real function on [0, ). The transform
f(s) :=
_
0
e
sx
f(x) dx (s IR)
is called the Laplace transform of f.
If X is an absolutely continuous positive random variable with density f then
f(s) =
M
X
(s). The Laplace transform has the following properties.
Lemma 4.11.
i) If f(x) 0 a.e.
f(s
1
)
f(s
2
) s
1
s
2
.
ii)
[f[(s
1
) < =
[f[(s) < for all s s
1
.
iii)
f
t
(s) =
_
0
e
sx
f
t
(x) dx = s
f(s) f(0)
provided f
t
(x) exists a.e. and
[f[(s) < .
iv) lim
s
s
f(s) = lim
x0
f(x)
provided f
t
(x) exists a.e., lim
x0
f(x) exists and
[f[(s) < for an s large
enough.
80 4. THE CRAM
ER-LUNDBERG MODEL
v) lim
s0
s
f(s) = lim
x
f(x)
provided f
t
(x) exists a.e., lim
x
f(x) exists and
[f[(s) < for all s > 0.
vi)
f(s) = g(s) on (s
0
, s
1
) = f(x) = g(x) Lebesgue a.e. x [0, ).
t
(u)e
su
du =
_
0
(u)e
su
du
_
0
_
u
0
(u y) dG(y)e
su
du.
We have to determine the last integral.
_
0
_
u
0
(u y) dG(y)e
su
du =
_
0
_
y
(u y)e
su
du dG(y)
=
_
0
_
0
(u)e
s(u+y)
du dG(y) =
(s)M
Y
(s) .
Thus we get the equation
c(s
(s) (0)) =
(s)(1 M
Y
(s))
which has the solution
(s) =
c(0)
cs (1 M
Y
(s))
=
c
cs (1 M
Y
(s))
.
The Laplace transform of can easily be found as
(s) =
_
0
(1 (u))e
su
du =
1
s
(s) .
Example 4.2 (continued). For exponentially distributed claims we obtain
(s) =
c /
cs (1 /( +s))
=
c /
s(c /( +s))
=
(c /)( +s)
s(c( +s) )
.
and
(s) =
1
s
(c /)( +s)
s(c( +s) )
=
c( +s) (c /)( +s)
s(c( +s) )
=
1
c( +s)
=
c
1
/c +s
.
By comparison with the moment generating function of the exponential distribution
we recognize that
(u) =
c
e
(/c)u
.
4. THE CRAM
ER-LUNDBERG MODEL 81
Example 4.5 (continued). For the Laplace transform of we get
(s) =
c p/ (1 p)/
cs (1 p/( +s) (1 p)/( +s))
=
c p/ (1 p)/
s(c p/( +s) (1 p)/( +s))
and for the Laplace transform of
(s) =
c p/( +s) (1 p)/( +s) c +p/ +(1 p)/
s(c p/( +s) (1 p)/( +s))
=
sp/(( +s)) +s(1 p)/(( +s))
s(c p/( +s) (1 p)/( +s))
=
p( +s)/ + (1 p)( +s)/
c( +s)( +s) p( +s) (1 p)( +s)
=
p( +s)/ + (1 p)( +s)/
cs
2
+ (( +)c )s +c ((1 p) +p)
.
The denominator can be written as c(s + R)(s +
R) where R and
R are the two
solutions to (4.6) with R <
R. The Laplace transform of can be written in the
form
(s) =
A
R +s
+
B
R +s
for some constants A and B. Hence
(u) = Ae
Ru
+Be
Ru
.
A must be the constant appearing in the Cramer-Lundberg approximation. The
constant B can be found from (0). We can see that in this case the Cramer-
Lundberg approximation is not exact.
Recall that
1
is the rst ladder epoch. On the set
1
< the random variable
supu C
t
: t 0 is absolutely continuous with distribution function
II P[supu C
t
: t 0 x [
1
< ] = 1
c
(x) .
Let Z be a random variable with the above distribution. Its moment generating
function is
M
Z
(r) =
_
0
e
ru
c
t
(u) du =
c
_
r
c
cr (1 M
Y
(r))
_
1
c
__
= 1
c
+
c(c )
r
cr (M
Y
(r) 1)
.
82 4. THE CRAM
ER-LUNDBERG MODEL
We will later need the rst two moments of the above distribution function. Assume
that
2
< . The rst derivative of the moment generating function is
M
t
Z
(r) =
c(c )
cr (M
Y
(r) 1) r(c M
t
Y
(r))
(cr (M
Y
(r) 1))
2
=
c(c )
rM
t
Y
(r) (M
Y
(r) 1)
(cr (M
Y
(r) 1))
2
.
Note that
lim
r0
1
r
(cr (M
Y
(r) 1)) = c .
We nd
lim
r0
rM
t
Y
(r) (M
Y
(r) 1)
r
2
= lim
r0
M
t
Y
(r) +rM
tt
Y
(r) M
t
Y
(r)
2r
=
2
2
and thus
II E[Z] =
c(c )
2
2(c )
2
=
c
2
2(c )
. (4.12)
Assume now that
3
< . The second derivative of M
Z
(r) is
M
tt
Z
(r) =
c(c )
1
(cr (M
Y
(r) 1))
3
_
rM
tt
Y
(r)(cr (M
Y
(r) 1))
2(rM
t
Y
(r) (M
Y
(r) 1))(c M
t
Y
(r))
_
.
For the limit to 0 we nd
lim
r0
1
r
3
_
rM
tt
Y
(r)(cr (M
Y
(r) 1))
2(rM
t
Y
(r) (M
Y
(r) 1))(c M
t
Y
(r))
_
= lim
r0
1
3r
2
_
(M
tt
Y
(r) +rM
ttt
Y
(r))(cr (M
Y
(r) 1))
+rM
tt
Y
(r)(c M
t
Y
(r)) 2rM
tt
Y
(r)(c M
t
Y
(r))
+ 2(rM
t
Y
(r) (M
Y
(r) 1))M
tt
Y
(r)
_
=
3
3
(c ) +
2
lim
r0
rM
t
Y
(r) (M
Y
(r) 1)
r
2
=
3
3
(c ) +
2
2
2
.
Thus the second moment of Z becomes
II E[Z
2
] =
c(c )
3
(c )/3 +
2
2
/2
(c )
3
=
c
_
3
3(c )
+
2
2
2(c )
2
_
.
(4.13)
4. THE CRAM
ER-LUNDBERG MODEL 83
4.10. Approximations to
4.10.1. Diusion Approximations
Diusion approximations are based on the following
Proposition 4.12. Let C
(n)
t
be a sequence of Cramer-Lundberg processes with
initial capital u
(n)
= u, claim arrival intensities
(n)
= n, claim size distributions
G
(n)
(x) = G(x
n
_
(n)
(n)
= c + (
n 1).
Let =
_
0
y dG(y) and assume that
2
=
_
0
y
2
dG(y) < . Then
C
(n)
t
d
u +W
t
is a (c ,
2
)-Brownian motion.
Proof. See [55] or [45].
Intuitively we let the number of claims in a unit time interval go to innity and
make the claim sizes smaller in such a way that the distribution of C
(n)
1
u tends
to a normal distribution and II E[C
(n)
1
u] = c . Let
(n)
denote the ruin time
of C
(n)
t
and = inft 0 : u + W
t
< 0 the ruin probability of the Brownian
motion. Then
Proposition 4.13. Let (C
(n)
t
) and (W
t
) be as above. Then
lim
n
II P[
(n)
t] = II P[ t]
and
lim
n
II P[
(n)
< ] = II P[ < ] .
Proof. The result for a nite time horizon is a special case of [87, Thm.9], see
also [55] or [45]. The result for the innite time horizon can be found in [70].
The idea of the diusion approximation is to approximate II P[
(1)
t] by II P[ t]
and II P[
(1)
< ] by II P[ < ]. Thus we need the ruin probabilities of the Brownian
motion.
84 4. THE CRAM
ER-LUNDBERG MODEL
Lemma 4.14. Let W
t
be a (m,
2
)-Brownian motion with m > 0 and =
inft 0 : u +W
t
< 0. Then
II P[ < ] = e
2um/
2
and
II P[ t] = 1
_
mt +u
t
_
+ e
2um/
2
_
mt u
t
_
.
Proof. By Lemma D.3 the process
exp
_
2m(u +W
t
)
2
_
is a martingale. By the stopping theorem
exp
_
2m(u +W
t
)
2
_
is a positive bounded martingale. Thus, because lim
t
W
t
= ,
exp
_
2um
2
_
= II E
_
exp
_
2m(u +W
2
__
= II P[ < ] .
It is easy to see that sW
1/s
m is a (0,
2
)-Brownian motion (see [58, p.351]) and
thus s(u + W
1/s
) m is (u,
2
)-Brownian motion. Denote the latter process by
W
s
. Then
II P[ t] = II P[infu +W
s
: 0 < s t < 0]
= II P[infs(u +W
1/s
) : s 1/t < 0]
= II E[II P[infm+
W
s
: s 1/t < 0 [
W
1/t
]]
=
_
m
1
_
2
2
/t
e
(yu/t)
2
2
2
/t
dy +
_
m
e
2u(y+m)
2
1
_
2
2
/t
e
(yu/t)
2
2
2
/t
dy
=
_
mu/t
/
t
_
+ e
2um
2
_
m
1
_
2
2
/t
e
(y+u/t)
2
2
2
/t
dy
= 1
_
mt +u
t
_
+ e
2um
_
mt u
t
_
.
Diusion approximations only work well if c/() is close to 1. There also exist
corrected diusion approximations which work much better, see [77] or [7].
4. THE CRAM
ER-LUNDBERG MODEL 85
u 0 0.25 0.5 0.75 1
(u) 0.6111 0.5246 0.4547 0.3969 0.3479
DA 1.0000 0.8071 0.6514 0.5258 0.4244
Er 63.64 53.87 43.26 32.49 21.98
u 1.25 1.5 1.75 2 2.25
(u) 0.3059 0.2696 0.2379 0.2102 0.1858
DA 0.3425 0.2765 0.2231 0.1801 0.1454
Er 11.96 2.54 -6.22 -14.32 -21.78
Table 4.2: Diusion approximation to ruin probabilities
Example 4.7 (continued). Let c = = 1 and G(x) = 1
1
3
(e
x
+ e
2x
+ e
3x
).
We nd c = 7/18 and
2
= 49/54. This leads to the diusion approxi-
mation (u) exp6u/7. Table 4.2 shows exact values ((u)), the diusion
approximation (DA) and the relative error multiplied by 100 (Er). Here we have
c/() = 18/11 = 1.63636 is not close to one. This is also indicated by the gures.
C
t
where
C
t
has exponentially distributed claim amounts and
II E[(C
t
u)
k
] = II E[(
C
t
u)
k
] for k = 1, 2, 3.
The rst three (centralized) moments are
II E[C
t
u] = (c )t =
_
c
_
t ,
Var[C
t
] = Var[u +ct C
t
] =
2
t =
2
2
t
and
II E[(C
t
II E[C
t
])
3
] = II E[(u +ct C
t
II E[u +ct C
t
])
3
] =
3
t =
6
3
t .
The parameters of the approximation are
=
3
2
3
,
86 4. THE CRAM
ER-LUNDBERG MODEL
u 0 0.25 0.5 0.75 1
(u) 0.6111 0.5246 0.4547 0.3969 0.3479
DV 0.5774 0.5102 0.4509 0.3984 0.3520
Er -5.51 -2.73 -0.86 0.38 1.18
u 1.25 1.5 1.75 2 2.25
(u) 0.3059 0.2696 0.2379 0.2102 0.1858
DV 0.3110 0.2748 0.2429 0.2146 0.1896
Er 1.67 1.95 2.07 2.09 2.03
Table 4.3: DeVylder approximation to ruin probabilities
=
2
2
2
=
9
3
2
2
2
3
and
c = c +
= c +
3
2
2
2
3
.
Thus the approximation to the probability of ultimate ruin is
(u)
c
e
u
.
There is also a formula for the probability of ruin within nite time. Let =
_
/( c). Then
(u, t)
c
e
(
/ c)u
_
0
f(x) dx (4.14)
where
f(x) =
exp2 ct cos x ( c +
)t + u( cos x 1)
1 +
2
2 cos x
(cos( u sin x) cos( u sin x + 2x)) .
A numerical investigation shows that the approximation is quite accurate.
Example 4.7 (continued). In addition to the previously calculated values we also
need
3
= 251/108. The approximation parameters are = 1.17131,
= 0.622472
and c = 0.920319. This leads to the approximation (u) 0.577441e
0.494949u
. It
turns out that the approximation works well.
4. THE CRAM
ER-LUNDBERG MODEL 87
4.10.3. The Beekman-Bowers Approximation
Recall from (4.12) and (4.13) that
F(u) = 1
c
(u)
is a distribution function and that
_
0
z dF(z) =
c
2
2(c )
and that
_
0
z
2
dF(z) =
c
_
3
3(c )
+
2
2
2(c )
2
_
.
The idea is to approximate the distribution function F by the distribution function
F(u) of a (, ) distributed random variable such that the rst two moments
coincide. Thus the parameters and have to full
=
c
2
2(c )
,
( + 1)
2
=
c
_
3
3(c )
+
2
2
2(c )
2
_
.
The Beekman-Bowers approximation to the ruin probability is
(u) =
c
(1 F(u))
c
(1
F(u)) .
Remark. If 2 IIN then 2Z
2
2
is
2
distributed.
Example 4.7 (continued). For the Beekman-Bowers approximation we have to
solve the equations
= 1.90909 ,
( + 1)
2
= 7.71429 ,
which yields the parameters = 0.895561 and = 0.469104. From this the Beek-
man Bowers approximation can be obtained. Here we have 2 = 1.79112 which is
not close to an integer. Anyway, one can interpolate between the
2
1
and the
2
2
distribution function to get the approximation
0.20888
2
1
(2u) + 0.79112
2
2
(2u)
to 1 c/() (u). Table 4.4 shows the exact values ((u)), the Beekman-Bowers
approximation (BB1) and the approximation obtained by interpolating the
2
dis-
tributions (BB2). The relative errors (Er) are given in percent. One can clearly see
that all the approximations work well.
88 4. THE CRAM
ER-LUNDBERG MODEL
u 0 0.25 0.5 0.75 1
(u) 0.6111 0.5246 0.4547 0.3969 0.3479
BB1 0.6111 0.5227 0.4553 0.3985 0.3498
Er 0.00 -0.35 0.12 0.42 0.54
BB2 0.6111 0.5105 0.4456 0.3914 0.3450
Er 0.00 -2.68 -2.02 -1.38 -0.83
u 1.25 1.5 1.75 2 2.25
(u) 0.3059 0.2696 0.2379 0.2102 0.1858
BB1 0.3076 0.2709 0.2387 0.2106 0.1859
Er 0.54 0.47 0.34 0.19 0.04
BB2 0.3046 0.2693 0.2383 0.2110 0.1869
Er -0.42 -0.11 0.18 0.40 0.59
Table 4.4: Beekman-Bowers approximation to ruin probabilities
4.11. Subexponential Claim Size Distributions
Let us now consider subexponential claim size distributions. In this case the Lund-
berg exponent does not exist (Lemma F.3).
Theorem 4.15. Assume that the ladder height distribution
1
_
x
0
(1 G(y)) dy
is subexponential. Then
lim
u
(u)
_
u
(1 G(y)) dy
=
c
.
Remark. Recall that the probability that ruin occurs at the rst ladder time
given there is a rst ladder epoch is
1
_
u
(1 G(y)) dy .
Hence the ruin probability is asymptotically (c )
1
times the probability of
ruin at the rst ladder time given there is a rst ladder epoch. But (c )
1
is
the expected number of ladder times. Intuitively for u large ruin will occur if one of
the ladder heights is larger than u.
4. THE CRAM
ER-LUNDBERG MODEL 89
Proof. Let B(x) denote the distribution function of the rst ladder height L
1
, i.e.
B(x) =
1
_
x
0
(1 G(y)) dy .
Choose > 0 such that (1 +) < c. By Lemma F.6 there exists D such that
1 B
n
(x)
1 B(x)
D(1 +)
n
.
From the Pollaczek-Khintchine formula (4.11) we obtain
(u)
1 B(u)
=
_
1
c
_
n=1
_
c
_
n
1 B
n
(u)
1 B(u)
D
_
1
c
_
n=1
_
c
_
n
(1 +)
n
< .
Thus we can interchange sum and limit. Recall from Lemma F.7 that
lim
u
1 B
n
(u)
1 B(u)
= n.
Thus
lim
u
(u)
1 B(u)
=
_
1
c
_
n=1
n
_
c
_
n
=
_
1
c
_
n=1
n
m=1
_
c
_
n
=
_
1
c
_
m=1
n=m
_
c
_
n
=
m=1
_
c
_
m
=
/c
1 /c
=
c
.
dy =
1
_
+x
_
1
we obtain
(u)
/( 1)
c /( 1)
_
+u
_
1
=
c( 1)
_
+u
_
1
.
90 4. THE CRAM
ER-LUNDBERG MODEL
u (u) App Er
1 0.364 8.79 10
3
-97.588
2 0.150 1.52 10
4
-99.898
3 6.18 10
2
8.58 10
6
-99.986
4 2.55 10
2
9.22 10
7
-99.996
5 1.05 10
2
1.49 10
7
-99.999
10 1.24 10
4
3.47 10
10
-100
20 1.75 10
8
5.40 10
13
-99.997
30 2.50 10
12
1.10 10
14
-99.56
40 1.60 10
15
6.71 10
16
-58.17
50 1.21 10
16
7.56 10
17
-37.69
Table 4.5: Approximations for subexponential claim sizes
Choose now c = 1, = 9, = 11 and = 1. Table 4.5 gives the exact value
((u)), the approximation (App) and the relative error in percent (Er). Consider
for instance u = 20. The ruin probability is so small, that it is not interesting for
practical purposes anymore. But the approximation still underestimates the true
value by almost 100%. That means we are still not far out enough in the tail. This
is the problem by using the approximation. It should, however, be remarked, that
for small values of the approximation works much better. Values (1, 2) are
also more interesting from a practical point of view.
Remark. The conditions of the theorem are also fullled for LN(,
2
), for
LG(, ) and for Wei(, c) ( < 1) distributed claims, see [34] and [60].
4.12. The Time of Ruin
Consider the function
f
(u) = II E[e
1I
<
[ C
0
= u] .
The function is dened at least for 0. We will rst nd a dierential equation
for f
(u).
Lemma 4.17. The function f
(u) +
_
_
u
0
f
(u y) dG(y) + 1 G(u) f
(u)
_
f
(u) = 0 . (4.15)
4. THE CRAM
ER-LUNDBERG MODEL 91
Proof. For h small we get
f
(u) = e
h
e
h
f
(u +ch)
+
_
h
0
_
_
u+ct
0
e
t
f
(u +ct y) dG(y) + e
t
(1 G(u +ct))
_
e
t
dt .
We see that f
(u +ch) f
(u)
ch
1 e
(+)h
h
f
(u +ch)
+
1
h
_
h
0
_
_
u+ct
0
f
(s) =
_
0
e
su
f
(u)e
su
du = s
(s) f
(0) ,
_
0
_
u
0
f
(u y) dG(y)e
su
du =
f
(s)M
Y
(s)
and
_
0
_
u
dG(y)e
su
du =
_
0
_
y
0
e
su
du dG(y) =
1 M
Y
(s)
s
.
Multiplying (4.15) by e
su
and integrating yields
c(s
(s) f
(0)) +
_
(s)M
Y
(s) +
1 M
Y
(s)
s
f
(s)
_
(s) = 0 .
Solving for
f
(s) yields
(s) =
cf
(0) s
1
(1 M
Y
(s))
cs (1 M
Y
(s))
. (4.16)
We know that
f
ER-LUNDBERG MODEL
is convex, has value < 0 at 0 and converges to as s . Thus there exists a
strictly positive root s() of the denominator. Because
f
(0) = s()
1
(1 M
Y
(s())) .
The function s() is dierentiable by the implicit function theorem
s
t
()(c M
t
Y
(s())) 1 = 0 . (4.17)
Because s(0) = 0 we obtain that lim
0
s() = 0.
Example 4.18. Let the claims be Exp() distributed. We have to solve
cs
s
+s
= 0
which admits the two solutions
s
() =
(c )
_
(c )
2
+ 4c
2c
where s
() < 0 s
+
(). Thus
(s) =
/( +s
+
()) /( +s)
cs s/( +s)
=
c( +s
+
())(s s
())
.
It follows that
f
(u) =
c( +s
+
())
e
s
() u
.
Noting that
II E[1I
<
] = lim
0
II E[e
1I
<
] = lim
0
d
d
II E[e
1I
<
]
we see
_
0
II E[1I
<
[ C
0
= u]e
su
du = lim
0
d
d
_
0
f
(u)e
su
du = lim
0
d
d
(s) .
We can nd the following explicit formula.
4. THE CRAM
ER-LUNDBERG MODEL 93
Lemma 4.19. Assume
2
< . Then
II E[1I
<
] =
1
c
_
2
2(c )
(u)
_
u
0
(u y)(y) dy
_
. (4.18)
Proof. We get
d
d
(s) =
s
t
()
cs (1 M
Y
(s))
1 M
Y
(s()) M
t
Y
(s())s()
s()
2
((1 M
Y
(s()))s()
1
(1 M
Y
(s))s
1
)
(cs (1 M
Y
(s)) )
2
. (4.19)
It follows from (4.17) that
s
t
(0) =
1
c
and we have already seen that
lim
0
1 M
Y
(s())
s()
= lim
s0
1 M
Y
(s)
s
= lim
s0
M
t
Y
(s) = .
Moreover,
lim
0
(1 M
Y
(s())) M
t
Y
(s())s()
s()
2
= lim
s0
(1 M
Y
(s)) M
t
Y
(s)s
s
2
= lim
s0
sM
tt
Y
(s)
2s
=
2
2
.
Letting tend to 0 in (4.19) yields
_
0
II E[1I
<
[ C
0
= u] e
su
du
=
2
2(c )(cs (1 M
Y
(s)))
s
1
(1 M
Y
(s))
(cs (1 M
Y
(s)))
2
=
2
2(c )
2
c
cs (1 M
Y
(s))
1
c
c
cs (1 M
Y
(s))
_
1
s
c
cs (1 M
Y
(s))
_
=
1
c
_
2
2(c )
(s)
(s)
(s)
_
.
But this is the Laplace transform of the assertion.
94 4. THE CRAM
ER-LUNDBERG MODEL
Corollary 4.20. Let t > 0. Then
II P[t < < ] <
2
2(c )
2
t
.
Proof. Because (u) and (u) take values in [0, 1] it is clear from (4.18) that
II E[1I
<
] <
2
2(c )
2
.
By Markovs inequality
II P[t < < ] = II P[1I
<
> t] <
1
t
II E[1I
<
] <
2
2(c )
2
t
.
_
u
0
c
e
R(uy)
_
1
c
e
Ry
_
dy
_
=
c
_
(c )
_
1
c
e
Ru
_
c
e
Ru
_
c
c
(e
Ru
1)
c
u
__
=
c
2
(c )
e
Ru
(u +c) =
1
c(c )
(u)(u +c) .
The conditional expectation of the time of ruin is linear in u
II E[ [ < ] =
u +c
c(c )
.
4. THE CRAM
ER-LUNDBERG MODEL 95
4.13. Seals Formulae
We consider now the probability of ruin within nite time (u, t). But rst we want
to nd the conditional nite ruin probability given C
t
for some t xed.
Lemma 4.21. Let t be xed, u = 0 and 0 < y ct. Then
II P[C
s
0, 0 s t [ C
t
= y] =
y
ct
.
Proof. Consider
C
t
C
(ts)
= cs
Nt
i=N
(ts)
+1
Y
i
.
This is also a Cramer-Lundberg model. Thus
II P[C
s
0, 0 s t [ C
t
= y] = II P[C
t
C
ts
0, 0 s t [ C
t
C
0
= y]
= II P[C
s
C
t
, 0 s t [ C
t
= y] .
Let S
s
= cs C
s
denote the aggregate claims up to time s. Denote by the
permutations of 1, 2, . . . , n. Then
II E[S
s
[ S
t
= y, N
t
= n, N
s
= k] =
1
n!
II E
_
i=1
Y
(i)
S
t
= y, N
t
= n, N
s
= k
_
=
k(n 1)!
n!
II E
_
n
i=1
Y
i
S
t
= y, N
t
= n, N
s
= k
_
=
ky
n
.
Because, given N
t
= n, the claim times are uniformly distributed in [0, t] (Proposi-
tion C.2) we obtain
II E[S
s
[ S
t
= y, N
t
= n] =
n
k=0
_
n
k
_
_
s
t
_
k
_
t s
t
_
nk
ky
n
=
n
k=1
_
n 1
k 1
_
_
s
t
_
k
_
t s
t
_
nk
y =
sy
t
.
This is independent of n. Thus
II E[S
s
[ S
t
= y] =
sy
t
and
II E[C
s
[ C
t
= y] = II E[cs S
s
[ S
t
= ct y] = cs
s(ct y)
t
=
sy
t
.
96 4. THE CRAM
ER-LUNDBERG MODEL
Then for v s < t
II E
_
y C
s
t s
C
v
, C
t
= y
_
=
y C
v
II E[(C
s
C
v
) [ C
v
, C
t
= y]
t s
=
y C
v
(s v)(y C
v
)/(t v)
t s
=
y C
v
t v
and the process
M
s
=
y C
s
t s
(0 s < t)
is a conditional martingale. Note that lim
st
M
s
= c. Let T = infs 0 : C
s
= y.
Because
0 M
s
1I
T>s
c
on C
t
= y we have
lim
st
II E[M
s
1I
T>s
[ C
t
= y] = II E
_
lim
st
M
s
1I
T>s
C
t
= y
_
= cII P[T = t [ C
t
= y] .
Note that by the stopping theorem M
Tt
is a bounded martingale. Thus because
M
T
= 0 on T < t
y
t
= M
0
= lim
st
II E[M
Ts
[ C
t
= y] = lim
st
II E[M
s
1I
T>s
[ C
t
= y] = cII P[T = t [ C
t
= y] .
Note that II P[T = t [ C
t
= y] = II P[C
s
C
t
, 0 s t [ C
t
= y].
Conditioning on C
t
is in fact a conditioning on the aggregate claim size S
t
.
Denote by F(x; t) the distribution function of S
t
. For the integration we use dF(; t)
for an integration with respect to the measure F(; t). Moreover, let (u, t) = 1
(u, t) = II P[ > t].
Theorem 4.22. For initial capital 0 we have
(0, t) =
1
ct
II E[C
t
0] =
1
ct
_
ct
0
F(y; t) dy .
Proof. By Lemma 4.21 it follows readily that
(0, t) = II E[II P[ > t [ C
t
]] = II E
_
C
t
0
ct
_
.
Moreover,
II E[C
t
0] = II E[(ct S
t
) 0] =
_
ct
0
(ct z) dF(z; t)
=
_
ct
0
_
ct
z
dy dF(z; t) =
_
ct
0
_
y
0
dF(z; t) dy =
_
ct
0
F(y; t) dy .
4. THE CRAM
ER-LUNDBERG MODEL 97
Let now the initial capital be arbitrary.
Theorem 4.23. With the notation used above we have for u > 0
(u, t) = F(u +ct; t)
_
u+ct
u
_
0, t
v u
c
_
F
_
dv;
v u
c
_
.
Proof. Note that
(u, t) = II P[C
t
> 0] II P[0 s < t : C
s
= 0, C
v
> 0 for s < v t]
and II P[C
t
> 0] = F(u + ct; t). Let T = sup0 s t : C
s
= 0 and set T = if
C
s
> 0 for all s [0, t]. Then, noting that II P[C
s
> 0 : 0 < s t [ C
0
= 0] = II P[C
s
0 : 0 < s t [ C
0
= 0],
II P[T [s, s + ds)] = II P[C
s
(c ds, 0], C
v
> 0 for v [s + ds, t]]
= (F(u +c(s + ds); s) F(u +cs; s))(0, t s ds)
= (0, t s)c dF(u +cs; s) .
Thus
(u, t) = F(u +ct; t)
_
t
0
(0, t s)c dF(u +cs; s)
= F(u +ct; t)
_
u+ct
u
_
0, t
v u
c
_
dF
_
v,
v u
c
_
.
> II E
_
e
rC
yu
(r)(yu)
; yu < yu
= II E
_
e
rC (r)
yu < yu
yu < yu
ER-LUNDBERG MODEL
Thus
II P[yu < yu [ C
0
= u] < e
(rmax(r)y,(r)y)u
.
Choosing the exponent as small as possible we obtain
II P[yu < yu [ C
0
= u] < e
R(y,y)u
(4.20)
where
R(y, y) = sup
rIR
minr (r)y, r (r)y .
The supremum over r IR makes sense because (r) > 0 for r < 0 and (r) =
if M
Y
(r) = . Since (R) = 0 it follows that R(y, y) R. Hence (4.20) could be
useful.
For further investigation of R(y, y) we consider the function f
y
(r) = r y(r).
Let r
= supr 0 : M
Y
(r) < . Then f
y
(r) exists in the interval (, r
)
and f
y
(r) = for r (r
, ). If M
Y
(r
) < then [f
y
(r
)[ < . Since
f
tt
y
(r) = yM
tt
Y
(r) < 0 the function f
y
(r) is concave. Thus there exists a unique r
y
such that f
y
(r
y
) = supf
y
(r) : r IR. Because f
y
(0) = 0 and f
t
y
(0) = 1 y
t
(0) =
1 +y(c ) > 0 we conclude that r
y
> 0. r
y
is either the solution to the equation
1 +y(c M
t
Y
(r
y
)) = 0
or r
y
= r
and let
y
0
= (M
t
Y
(R) c)
1
.
It follows r
y
0
= R. We call y
0
the critical value.
Because
d
dy
f
y
(r) = (r) it follows readily that
d
dy
f
y
(r) 0 r R.
We get
f
y
(r) f
y
(r) as r R.
Because M
t
Y
(r) is an increasing function it follows that
y y
0
r
y
R.
Thus
R(y, y) =
_
_
R if y y
0
y,
f
y
(r
y
) if y < y
0
,
f
y
(r
y
) if y > y
0
.
4. THE CRAM
ER-LUNDBERG MODEL 99
If y
0
[y, y] then we got again Lundbergs inequality. If y < y
0
then R(y, y) does
not depend on y. By choosing y as small as possible we obtain
II P[0 < yu [ C
0
= u] < e
R(0,y)u
(4.21)
for y < y
0
. Note that R(0, y) > R. Analogously
II P[yu < < [ C
0
= u] < e
R(y,)u
(4.22)
for y > y
0
. The strict inequality follows from
e
ru
II E[e
rC (r)
; yu < < ] > e
(r)yu
II P[yu < < ]
if 0 < r < R, compare also (4.5). Again R(y, ) > R.
We see that II P[ / ((y
0
)u, (y
0
+ )u)] goes faster to 0 than II P[ ((y
0
)u, (y
0
+)u)]. Intuitively ruin should therefore occur close to y
0
u.
Theorem 4.24. Assume that R < r
. Then
u
y
0
in probability on the set < as u .
Proof. Choose > 0. By the Cramer-Lundberg approximation II P[ < [ C
0
=
u] expRu C for some C > 0. Then
II P
_
u
y
0
>
< , C
0
= u
_
=
II P[ < (y
0
)u [ C
0
= u] + II P[(y
0
+)u < < [ C
0
= u]
II P[ < [ C
0
= u]
expR(0, y
0
)u + expR(y
0
+, )u
II P[ < [ C
0
= u]
=
exp(R(0, y
0
) R)u + exp(R(y
0
+, ) R)u
II P[ < [ C
0
= u]e
Ru
0
as u .
Bibliographical Remarks
The Cramer-Lundberg model was introduced by Filip Lundberg [61]. The basic
results, like Lundbergs inequality and the Cramer-Lundberg approximation go back
100 4. THE CRAM
ER-LUNDBERG MODEL
to Lundberg [62] and Cramer [24], see also [25], [46] or [66]. The dierential equation
for the ruin probability (4.2) and the Laplace transform of the ruin probability can
for instance be found in [25]. The renewal theory proof of the Cramer-Lundberg
approximation is due to Feller [40]. The martingale (4.4) was introduced by Gerber
[41]. The latter paper contains also the martingale proof of Lundbergs inequality
and the proof of (4.21). The proof of (4.22) goes back to Grandell [47]. Section 4.14
can also be found in [35]. Some similar results are obtained in [4] and [5]. Segerdahl
[76] proved the asymptotic normality of the ruin time conditioned on ruin occurs.
The results on reinsurance and ruin can be found in [86] and [22]. Proposition 4.9
can be found in [43, p.130].
The term severity of ruin was introduced in [44]. More results on the topic
are obtained in [33], [32], [72] and [66]. A diusion approximation, even though not
obtained mathematically correct, was already used by Hadwiger [51]. The modern
approach goes back to Iglehart [55], see also [45], [7] and [70]. The deVylder ap-
proximation was introduced in [31]. Formula (4.14) was found by Asmussen [7]. It
is however stated incorrectly in the paper. For the correct formula see [16] or [66].
The Beekman-Bowers approximation can be found in [17].
Theorem 4.15 was obtain by von Bahr [85] in the special case of Pareto dis-
tributed claim sizes and by Thorin and Wikstad [84] in the case of lognormally
distributed claims. The theorem in the present form is basically found in [34], see
also [38].
Section 4.12 is taken from [69]. Parts can also be found in [70] and [16]. (4.16)
was rst obtained in [25]. The expected value of the ruin time for exponentially
distributed claims can be found in [43, p.138]. Seals formulae were rst obtained
by Takacs [80, p.56]. Seal [74] and [75] introduced them to risk theory. The present
presentation follows [43].
5. THE RENEWAL RISK MODEL 101
5. The Renewal Risk Model
5.1. Denition of the Renewal Risk Model
The easiest generalisation of the classical risk model is to assume that the claim
number process is a renewal process. Then the risk process is not Markovian any-
more, because the distribution of the time of the next claim is dependent on the
past via the time of the last claim. We therefore cannot nd a dierential equation
for (u) as in the classical case.
Let
C
t
= u +ct
Nt
i=1
Y
i
where
N
t
is a renewal process with event times 0 = T
0
, T
1
, T
2
, . . . with interarrival
distribution function F, mean
1
and moment generating function M
T
(r). We
denote by T a generic random variable with distribution function F. The rst
claim time T
1
has distribution function F
1
. If nothing else is said we consider
the ordinary case where F
1
= F.
The claim amounts Y
i
build an iid. sequence with distribution function G,
moments
k
= II E[Y
k
i
] and moment generating function M
Y
(r). As in the classical
model =
1
.
N
t
and Y
i
are independent.
This model is called the renewal risk model or, after the person who introduced
the model, the Sparre Andersen model.
As in the classical model we dene the ruin time = inft > 0 : C
t
< 0 and
the ruin probabilities (u, t) = II P[ t] and (u) = II P[ < ]. Considering the
random walk C
T
i
we can conclude that
(u) = 1 II E[C
T
2
C
T
1
] = c/ 0 .
In order to avoid ruin a.s. we assume the net prot condition c > . Note that in
contrary to the classical case
II E[C
t
u] ,= (c )t for most t IR
102 5. THE RENEWAL RISK MODEL
except if T has an exponential distribution. But
lim
t
1
t
(C
t
u) = c
by the law of large numbers. It follows again that (u) 0 as u because
infC
t
u, t 0 is nite.
5.2. The Adjustment Coecient
We were able to construct some martingales of the form exprC
t
(r)t in
the classical case. Because the risk process C
t
is not Markov anymore we cannot
hope to get such a martingale again because Lemma 4.1 does not hold any more.
But the claim times are renewal times because C
T
1
+t
C
T
1
is a renewal risk model
independent of T
1
and Y
1
.
Lemma 5.1. Let C
t
be an (ordinary) renewal risk model. For any r IR such
that inf1/M
T
(s) : s 0, M
T
(s) < M
Y
(r) < let (r) be the unique solution
to
M
Y
(r)M
T
((r) cr) = 1 . (5.1)
Then the discrete time process
exprC
T
i
(r)T
i
is a martingale.
Remark. In the classical case with F(t) = 1 e
t
equation (5.1) is
M
Y
(r)
+(r) +cr
= 1
which is equivalent to the equation (4.3) in the classical case.
Proof. Note that 1/M
Y
(r) supM
T
(s) : M
T
(s) < . Because M
T
(r) is an
increasing and continuous function dened for all r 0 and because M
T
(r) 0 as
r there exists a unique solution (r) to (5.1). Then
II E[e
rC
T
i+1
(r)T
i+1
[ T
T
i
] = II E[e
r(c(T
i+1
T
i
)Y
i+1
)(r)(T
i+1
T
i
)
[ T
T
i
]e
rC
T
i
(r)T
i
= II E[e
rY
i+1
e
(cr+(r))(T
i+1
T
i
)
]e
rC
T
i
(r)T
i
= M
Y
(r)M
T
(cr (r))e
rC
T
i
(r)T
i
= e
rC
T
i
(r)T
i
.
t
(r) +c > 0. Thus
tt
(r) > 0 and (r) is a strictly convex function.
Let r = 0, i.e. (0) = 0. Then
t
(0) =
M
t
Y
(0)M
T
(0)
M
Y
(0)M
t
T
(0)
c = c < 0 (5.3)
by the net prot condition. Thus there exists at most a second solution to (R) = 0
with R > 0. Again we call this solution adjustment coecient or Lundberg
exponent. Note that R is the strictly positive solution to the equation
M
Y
(r)M
T
(cr) = 1 .
Example 5.2. Let C
t
be a renewal risk model with Exp(1) distributed claim
amounts, premium rate c = 5 and interarrival time distribution
F(t) = 1
1
2
(e
3t
+ e
7t
) .
It follows that M
Y
(r) exists for r < 1, M
T
(r) exists for r < 3 and = 4.2. The net
prot condition 5 > 4.2 is fullled. The equation to solve is
1
1 r
1
2
_
3
3 + 5r
+
7
7 + 5r
_
= 1 .
104 5. THE RENEWAL RISK MODEL
Thus
3(7 + 5r) + 7(3 + 5r) = 2(1 r)(3 + 5r)(7 + 5r)
or equivalently
25r
3
+ 25r
2
4r = 0 .
We nd the obvious solution r = 0 and the two other solutions
r
1/2
=
25
1025
50
=
5
41
10
=
_
0.140312,
1.14031.
From the theory we conclude that R = 0.140312. But we proved that there is no
negative solution. Why do we get r
2
= 1.14031? Obviously M
Y
(r
2
) < 1 < .
But cr
2
= 5.70156 > 3 and thus M
T
(cr
2
) = . Thus r
2
is not a solution to the
equation M
Y
(r)M
T
(cr) = 1.
Example 5.3. Let T (, ). Assume that there exists an r
such that
M
Y
(r) < r < r
. Then
lim
rr
M
Y
(r) = .
Let R(, ) denote the Lundberg exponent which exists in this case and (r; , )
denote the corresponding (r). Then for h > 0
_
+cR(, )
_
+h
M
Y
(R(, )) =
_
+cR(, )
_
h
< 1 .
That means that (R(, ); + h, ) < 0 and therefore R( + h, ) > R(, ).
Similarly
_
+h
+h +cR(, )
_
M
Y
(R(, )) >
_
+cR(, )
_
M
Y
(R(, )) = 1
and thus R(, + h) < R(, ). It would be interesting to see what happens if we
let 0 and 0 in such a way that the mean value remains constant. Let
therefore = . Recall that m
Y
(r) = log M
Y
(r).
Observe that
lim
0
log
_
+cr
_
+m
Y
(r) = m
Y
(r)
and therefore R() = R(, ) 0 as 0. The latter follows because conver-
gence to a continuous function is uniform on compact intervals. Let now r() =
1
R(). Then r() has to solve the equation
log
_
+cr
_
+
m
Y
(r)
= 0 . (5.4)
5. THE RENEWAL RISK MODEL 105
Letting 0 yields
log
_
+cr
_
+rm
t
Y
(0) = log
_
+cr
_
+r = 0 .
The latter function is convex, has the root r = 0, the derivative c/ + < 0 at
0 and converges to innity as r . Thus there exists exactly one additional
solution r
0
and r() r
0
as 0.
Replace r by r() in (5.4) and take the derivative, which exists by the implicit
function theorem,
cr
t
()
+cr()
+
(r
t
() +r())m
t
Y
(r())
m
Y
(r())
2
= 0 .
One obtains
r
t
() =
r()m
Y
(r())m
Y
(r())
2
c
+cr()
m
t
Y
(r())
.
Note that by Taylor expansion
r()m
t
Y
(r()) m
Y
(r()) =
2
2
2
r()
2
+O(
3
r()
3
) =
2
2
2
r
2
0
+O(
3
)
uniformly on compact intervals for which R() exists. The second equality holds
because lim
0
r
t
() = r
0
> 0 exists. We denoted by
2
the variance of the claims.
We found
r
t
() =
2
r
2
0
( +cr
0
)
2(c cr
0
)
+O() .
Because the convergence is uniform it follows that
r() = r
0
+
2
r
2
0
( +cr
0
)
2(c cr
0
)
+O(
2
)
or
R() = r
0
+
2
r
2
0
( +cr
0
)
2(c cr
0
)
2
+O(
3
) .
Note that the moment generating function of the interarrival times
_
r
_
=
_
1
r
e
r/
converges to the moment generating function of deterministic interarrival times T =
1/. That means that T() 1/ in distribution as 0. But the Lundberg
exponents do not converge to the Lundberg exponent of the model with deterministic
interarrival times.
106 5. THE RENEWAL RISK MODEL
5.3. Lundbergs Inequality
5.3.1. The Ordinary Case
In order to nd an upper bound for the ruin probability we try to proceed as in the
classical case. Unfortunately we cannot nd an analogue to (4.1). But recall the
interpretation of (4.2) as conditioning on the ladder heights. Let H(x) = II P[ <
, C
x [ C
0
= 0]. Then
(u) =
_
u
0
(u x) dH(x) + (H() H(u)) (5.5)
where H() = II P[ < [ C
0
= 0]. Unfortunately we cannot nd an explicit
expression for H(u). In order to nd a result we have to link the Lundberg exponent
R to H.
Lemma 5.4. Let C
t
be an ordinary renewal risk model with Lundberg exponent
R. Then
_
0
e
Rx
dH(x) = II E
_
e
RC
1I
<
C
0
= 0
= 1 .
Proof. Let B(x) be the weekly descending ladder height of the process C
Tn
.
Multiplying the Wiener-Hopf-factorization (E.1) by e
rx
and integrating yields
M
Y
(r)M
T
(cr) = M
H
(r) +M
B
(r) M
H
(r)M
B
(r) .
We see that M
H
(r) < i M
Y
(r) < and M
B
(r) < i M
T
(cr) < . For
r = R the left hand side is one and we can rewrite the equation as (M
H
(R)
1)(1 M
B
(R)) = 0. Since B(x) is distribution on the negative half line we have
M
B
(R) < 1. Thus M
H
(R) = 1, which is the assertion.
Also here we can give a martingale proof. The discrete time process e
RC
T
i
is
a martingale with initial value 1 if C
0
= 0. We will drop the conditioning on C
0
= 0
for the rest of the proof. By the stopping theorem
1 = II E
_
e
RC
Tn
= II E
_
e
RC
1I
Tn
+ II E
_
e
RC
Tn
1I
>Tn
.
The function e
RC
1I
Tn
is increasing with n. By the monotone limit theorem it
follows that
lim
n
II E[e
RC
1I
Tn
] = II E
_
lim
n
e
RC
1I
Tn
_
= II E[e
RC
1I
<
] .
5. THE RENEWAL RISK MODEL 107
For the second term we have e
RC
Tn
1I
>Tn
1 and we know that C
Tn
as
n . Thus
lim
n
II E[e
RC
Tn
1I
>Tn
] = II E
_
lim
n
e
RC
Tn
1I
>Tn
_
= 0 .
We are now ready to prove Lundbergs inequality.
Theorem 5.5. Let C
t
be an ordinary renewal risk model and assume that the
Lundberg exponent R exists. Then
(u) < e
Ru
.
Proof. Assume that the assertion were wrong and let
u
0
= infu 0 : (u) e
Ru
.
Let u
0
u < u
0
+ such that (u) e
Ru
. Then
(u
0
) (u) e
Ru
> e
R(u
0
+)
.
Thus (u
0
) e
Ru
0
. Because e
RC
> 1 it follows from Lemma 5.4 that (0) =
H() < 1 and therefore u
0
R
1
log H() > 0. Assume rst that H(u
0
) = 0.
Then it follows from (5.5) that
e
Ru
0
(u
0
) = (0)H(u
0
) +
_
u
0
dH(x) < H(u
0
) +
_
u
0
e
R(xu
0
)
dH(x)
=
_
u
0
e
R(xu
0
)
dH(x) = e
Ru
0
_
0
e
Rx
dH(x) = e
Ru
0
which is a contradiction. Thus H(u
0
) > 0.
As in the classical case we obtain from (5.5) and Lemma 5.4
e
Ru
0
(u
0
) =
_
u
0
0
(u
0
x) dH(x) +
_
u
0
dH(x)
<
_
u
0
0
e
R(u
0
x)
dH(x) +
_
u
0
dH(x)
_
0
e
R(u
0
x)
dH(x) = e
Ru
0
.
This is a contradiction and the theorem is proved.
108 5. THE RENEWAL RISK MODEL
As in the classical case we prove the result using martingale techniques. By the
stopping theorem
e
Ru
= e
RC
T
0
= II E
_
e
RC
Tn
II E
_
e
RC
Tn
; T
n
= II E
_
e
RC
; T
n
.
Letting n yields by monotone convergence
e
Ru
II E
_
e
RC
; <
> II P[ < ]
because C
< 0.
Example 5.6. Let Y
i
be Exp() distributed and
F(t) = 1 pe
t
(1 p)e
t
where 0 < < and 0 < p ( )
1
. The net prot condition can then be
written as
c( +p( )) > .
The equation for the adjustment coecient is
r
_
p
+cr
+
(1 p)
+cr
_
= 1 .
This is equivalent to
c
2
r
3
c(c )r
2
(c( +p( )) )r = 0
from which the solutions r
0
= 0 and
r
1/2
=
c
_
(c +)
2
+ 4cp( )
2c
follow. Note that by the net prot condition r
1
> 0 > r
2
. By the condition on p
r
2
< r
1
c +
_
(c +)
2
+ 4c
2c
<
c +
_
(c + +)
2
2c
= .
Thus M
Y
(r
1/2
) < . And
cr
2
=
+ c +
_
(c +)
2
+ 4cp( )
2
>
+ c +
_
(c +)
2
2
5. THE RENEWAL RISK MODEL 109
and thus M
T
(cr
2
) = , i.e. r
2
is not a solution. Therefore
R =
c +
_
(c +)
2
+ 4cp( )
2c
.
It follows from Lundbergs inequality that
(u) < e
Ru
.
o
(u). We know that C
T
1
+t
is an ordinary model with initial capital C
T
1
. There
are two possibilities: Ruin occurs at T
1
or C
T
1
0. Thus
(u) =
_
0
_
_
u+ct
0
o
(u +ct y) dG(y) +
_
u+ct
dG(y)
_
dF
1
(t)
<
_
0
_
_
u+ct
0
e
R(u+cty)
dG(y) +
_
u+ct
e
R(u+cty)
dG(y)
_
dF
1
(t)
=
_
0
_
0
e
R(u+cty)
dG(y) dF
1
(t)
= e
Ru
II E
_
e
R(Y
1
cT
1
)
= M
Y
(R)M
T
1
(cR)e
Ru
.
Thus (u) < Ce
Ru
for C = M
Y
(R)M
T
1
(cR). In the cases considered so far we
always had C = 1. But in the general case C (0, M
Y
(R)) can be any value of
this interval (let T
1
be deterministic). If F
1
(t) = F(t) then C = 1 by the equation
determining the adjustment coecient.
5.4. The Cramer-Lundberg Approximation
5.4.1. The Ordinary Case
In order to obtain the Cramer-Lundberg approximation we proceed as in the classical
case and multiply (5.5) by e
Ru
.
(u)e
Ru
=
_
u
0
(u x)e
R(ux)
e
Rx
dH(x) + e
Ru
(H() H(u)) .
We obtain the following
110 5. THE RENEWAL RISK MODEL
Theorem 5.7. Let C
t
be an ordinary renewal risk model. Assume that R exists
and that there exists an r > R such that M
Y
(r) < . If the distribution of Y
1
cT
1
given Y
1
cT
1
> 0 is not arithmetic then
lim
u
(u)e
Ru
=
1 H()
R
_
0
xe
Rx
dH(x)
=: C .
If the distribution of Y
1
cT
1
is arithmetic with span then for x [0, )
lim
n
(x +n)e
R(x+n)
= Ce
Rx
1 e
R
R
.
Proof. We prove only the non-arithmetic case. The arithmetic case can be proved
similarly. Recall from Lemma 5.4 that
_
x
0
e
Ry
dH(y) is a proper distribution func-
tion. It follows as in the classical case that e
Ru
(H() H(u)) is directly Riemann
integrable. From M
Y
(r) < for an r > R it follows that
_
0
xe
Rx
dH(x) < .
Thus by the renewal theorem
lim
u
(u)e
Ru
=
_
0
e
Ru
(H() H(u)) du
_
0
xe
Rx
dH(x)
.
Simplifying the numerator we nd
_
0
e
Ru
_
u
dH(x) du =
_
0
_
x
0
e
Ru
du dH(x) =
1
R
(1 H())
where we used Lemma 5.4 in the last equality.
It should be remarked that in general there is no explicit expression for the ladder
height distribution H(x). It is therefore dicult to nd an explicit expression for
C. Nevertheless we know that 0 < C 1.
Assume that we know C. Then
(u) Ce
Ru
for u large.
Example 5.8. Assume that Y
i
is Exp() distributed. Because M
Y
(r) as
r it follows that R exists and that the conditions of Theorem 5.7 are fullled.
Consider the martingale e
RC
T
i
. By the stopping theorem
e
Ru
= II E
_
e
RC
T
i
= II E
_
e
RC
1I
T
i
+ II E
_
e
RC
T
i
1I
>T
i
.
5. THE RENEWAL RISK MODEL 111
As in the proof of Lemma 5.4 it follows that
e
Ru
= II E
_
e
RC
1I
<
= II E
_
e
RC
<
(u) .
Assume for the moment that we know C
. Let Z = C
because
ruin occurs at time . Then
II P[C
> x [ C
= y, < ]
= II P[Y
1
> y +x [ Y
1
> y] = e
x
.
It follows that
II E
_
e
RC
[ <
=
_
0
e
Rx
e
x
dx =
R
.
Thus we obtain the explicit solution
(u) =
R
e
Ru
to the ruin problem. Therefore, as in the classical case, the Cramer-Lundberg ap-
proximation is exact and we have an explicit expression for C.
5.4.2. The general case
Consider the equation
(u) =
_
0
_
_
u+ct
0
o
(u +ct y) dG(y) + (1 G(u +ct))
_
dF
1
(t) .
We have to multiply the above equation by e
Ru
. We obtain by Lundbergs inequality
_
0
_
u+ct
0
o
(u +ct y)e
Ru
dG(y) dF
1
(t)
=
_
0
_
u+ct
0
o
(u +ct y)e
R(u+cty)
e
Ry
dG(y)e
cRt
dF
1
(t)
<
_
0
_
u+ct
0
e
Ry
dG(y)e
cRt
dF
1
(t) (5.6)
_
0
_
0
e
Ry
dG(y)e
cRt
dF
1
(t) = M
Y
(R)M
T
1
(cR) < .
We can therefore interchange limit and integration
lim
u
_
0
_
u+ct
0
o
(u +ct y)e
R(u+cty)
e
Ry
dG(y)e
cRt
dF
1
(t)
=
_
0
_
0
Ce
Ry
dG(y)e
cRt
dF
1
(t) = M
Y
(R)M
T
1
(cR)C .
112 5. THE RENEWAL RISK MODEL
The remaining term is converging to 0 because
_
0
_
u+ct
e
Ru
dG(y) dF
1
(t)
_
0
_
u+ct
e
R(yct)
dG(y) dF
1
(t)
_
0
_
u
e
Ry
dG(y) e
cRt
dF
1
(t)
= M
T
1
(cR)
_
u
e
Ry
dG(y) .
Hence
lim
u
(u)e
Ru
= M
Y
(R)M
T
1
(cR)C (5.7)
where C was the constant obtained in the ordinary case.
Example 5.8 (continued). In the general case we have to change the martingale
because in general II E[e
RC
T
1
] ,= e
Ru
. Note that
II E[e
RC
T
1
] = e
Ru
M
Y
(R)M
T
1
(cR) =
R
e
Ru
M
T
1
(cR) .
Let
M
n
=
_
e
RC
Tn
if n 1,
R
e
Ru
M
T
1
(cR) if n = 0.
As in the ordinary case we obtain
R
e
Ru
M
T
1
(cR) =
R
(u)
or equivalently
(u) = M
T
1
(cR)e
Ru
.
n
_
(n)
(n)
= c + (
n 1)
5. THE RENEWAL RISK MODEL 113
where =
(1)
and =
(1)
. Then
C
(n)
t
u +W
t
0
(1 U(x)) dx. Let G
1
(x) =
1
_
x
0
(1 G(y)) dy and U
1
(x) = m
1
_
x
0
(1
U(y)) dy. Then
i) If G
1
is subexponential, then U
1
is subexponential and
lim
x
1 U
1
(x)
1 G
1
(x)
=
m
.
114 5. THE RENEWAL RISK MODEL
ii) If U
1
is subexponential then H/H() is subexponential and
lim
x
1 U
1
(x)
H() H(x)
=
c
m(1 H())
.
Proof.
i) Using Fubinis theorem
m(1 U
1
(x)) =
_
x
_
0
(1 G(y +ct)) dF(t) dy
=
_
0
_
x
(1 G(y +ct)) dy dF(t)
=
_
0
(1 G
1
(x +ct)) dF(t) .
It follows by the bounded convergence theorem and Lemma F.2 that
lim
x
1 U
1
(x)
1 G
1
(x)
= lim
x
m
_
0
1 G
1
(x +ct)
1 G
1
(x)
dF(t) =
m
.
By Lemma F.8 U
1
is subexponential.
ii) Consider the random walk
n
i=1
(Y
i
cT
i
). We use the notation of Lemma E.2.
By the Wiener-Hopf factorisation for y > 0
1 U(y) =
_
0
_
b
x
(H(y z) H(y)) dy d(z)
=
_
0
_
_
xz
x
(H() H(y)) dy
_
bz
b
(H() H(y)) dy
_
d(z) .
Because
_
0
_
xz
x
(H() H(y)) dy d(z) .
We nd
_
0
[z[ d(z) .
5. THE RENEWAL RISK MODEL 115
For s > 0 we obtain
(H()H(x+s))
_
0
s
[z[ d(z)
_
0
s
[z[(H()H(xz)) d(z) m(1U
1
(x))
and therefore
1
_
0
[z[ d(z)
H() H(x +s)
m(1 U
1
(x +s))
_
0
[z[ d(z)
_
0
s
[z[ d(z)
1 U
1
(x)
1 U
1
(x +s)
.
Thus by Lemma F.2 (H() H(x))
1
(1 U
1
(x)) m
1
_
0
n
i=1
(Y
i
cT
i
).
u
(1 G(y)) dy
=
c
.
Remark. Note that this is exactly Theorem 4.15. The dierence is only that
G
1
(x) is not the ladder height distribution anymore.
Proof. Let us rst consider an ordinary renewal model. By repeating the proof
of Theorem 4.15 we nd that
lim
u
(u)
1 H(u)/H()
=
H()
1 H()
because H(x)/H() is subexponential by Proposition 5.11. Moreover, by Proposi-
tion 5.11
lim
u
(u)
_
u
(1 G(y)) dy
= lim
u
(u)
1 H(u)/H()
1
H()
H() H(u)
1 U
1
(u)
1 U
1
(u)
(1 G
1
(u))
=
H()
1 H()
1
H()
m(1 H())
c
1
m
=
c
.
For an arbitrary renewal risk model the assertion follows similarly as in Section 5.4.2.
But two properties of subexponential distributions that are not proved here will be
needed.
116 5. THE RENEWAL RISK MODEL
5.7. Finite Time Lundberg Inequalities
Let 0 y < y < and dene T = infT
n
: n IIN, T
n
yu. Using the stopping
theorem
e
ru
= II E
_
e
rC
TTn
(r)(TTn)
= II E
_
e
rC
T
(r)(T)
; T
n
T
+ II E
_
e
rC
Tn
(r)Tn
; T
n
< T
.
By monotone convergence
lim
n
II E
_
e
rC
T
(r)(T)
; T
n
T
= II E
_
e
rC
T
(r)(T)
.
For the second term we obtain
lim
n
II E
_
e
rC
Tn
(r)Tn
; T
n
< T
lim
n
maxe
(r)yu
, 1II P[T
n
< T] = 0 .
Therefore
e
ru
= II E
_
e
rC
T
(r)(T)
> II E
_
e
rC
T
(r)(T)
; yu < yu
= II E
_
e
rC (r)
yu < yu
yu < yu
= supr 0 : M
Y
(r) < > R. Let
y
0
= (
t
(R))
1
=
_
M
t
Y
(R)M
T
(cR)
M
Y
(R)M
t
T
(cR)
c
_
1
denote the critical value. Then we obtain as in the classical case
II P[0 < yu [ C
0
= u] < e
R(0,y)u
and R(0, y) > R if y < y
0
, and
II P[yu < < [ C
0
= u] < e
R(y,)u
and R(y, ) > R if y > y
0
. We can prove in the same way as Theorem 4.24 was
proved.
5. THE RENEWAL RISK MODEL 117
Theorem 5.13. Assume that R < r
. Then
u
y
0
in probability on the set < .
Bibliographical Remarks
The renewal risk model was introduced by E. Sparre Andersen [3]. A good reference
are also the papers by Thorin [81], [82] and [83]. Example 5.3 is due to Grandell [46,
p.71]. Lundbergs inequality in the ordinary case was rst proved by Andersen [3,
p.224]. Example 5.6 can be found in [46, p.60]. The Cramer-Lundberg approxima-
tion was rst proved by Thorin, [81, p.94] in the ordinary case and [82, p.97] in the
stationary case. The results on subexponential claim sizes go back to Embrechts
and Veraverbeke [38]. The results on nite time Lundberg inequalities can be found
in [35]. More results on the renewal risk model is to be found in Chapter 6 of [66].
118 6. SOME ASPECTS OF HEAVY TAILS
6. Some Aspects of Heavy Tails
6.1. Why Heavy Tailed Distributions Are Dangerous
In most of the considerations so far only the rst or the rst two moments of the
claim size distributions were needed. And mean value and variance are easy to
estimate via the estimators
=
1
n
n
i=1
Y
i
and
2
=
1
n 1
n
i=1
(Y
i
)
2
.
These estimators are unbiased under any underlying claim size distribution. Nev-
ertheless, we will see that using the estimator is dangerous in the heavy claim
case.
For the modelling of large claims the Pareto distribution is very popular. Let
Y
i
: 1 i n be a sample of iid. Pa(,1) distributed random variables where
> 1. Note that the second parameter is only a scale parameter. Thus it is no loss
of generality to choose it equal to 1. Assume that we got the knowledge that
Y
max
n
:= max
in
Y
i
M
where M is a constant. What is the expected value of conditioned on Y
max
n
M
and what is the relative error compared to the true mean ( 1)
1
?
For the conditional mean value we get
II E[ [ Y
max
n
M] = II E[ [ Y
1
M, . . . , Y
n
M] = II E[Y
1
[ Y
1
M, . . . , Y
n
M]
= II E[Y
1
[ Y
1
M] =
_
M
0
x
(1+x)
+1
dx
_
M
0
(1+x)
+1
dx
=
_
M
0
(x + 1)
(1+x)
+1
dx
_
M
0
(1+x)
+1
dx
1
=
1
1 (1 +M)
(1)
1 (1 +M)
1 .
The relative error is
II E[ [ Y
max
n
M] ( 1)
1
( 1)
1
=
(1 (1 +M)
(1)
)
1 (1 +M)
( 1) 1
=
(1 +M)
(1)
(1 +M)
1 (1 +M)
.
6. SOME ASPECTS OF HEAVY TAILS 119
1.5 2.0 2.5 3.0
0.5
0.4
0.3
0.2
0.1
Figure 6.1: Relative error for p = 0.99 and n = 1000.
Let us x p (0, 1) and let us choose M such that II P[Y
max
n
M] = p, i.e. we x
the probability that it is no mistake to assume that Y
max
n
M.
II P[Y
max
n
M] = II P[Y
1
M, . . . , Y
n
M] = II P[Y
1
M]
n
=
_
1 (1 +M)
_
n
= p
or equivalently
(1 +M)
1
=
_
1 p
1/n
_
1/
.
The relative error is therefore
(1 p
1/n
)
(1)/
(1 p
1/n
)
p
1/n
. (6.1)
Figure 6.1 shows the relative error made by choosing p = 0.99 and n = 1000.
n = 1000 is a quite large sample size (a typical example is re insurance). With
probability p = 0.99 the maximum value is below M, i.e. on a set with probability
0.99 the expected relative error is larger than (6.1).
Quite often estimation yields values for between 1 and 1.5. Some relative errors
may be found in the table below:
1.05 1.1 1.15 1.2 1.25 1.3 1.4 1.5
error -60.7% -38.6% -25.6% -17.6% -12.5% -9.1% -5.2% -3.2%
If for instance the true parameter is 1.25 then the premium computed would be
at least 12.5% too low in 99% of all cases. This is quite a crucial error for any
120 6. SOME ASPECTS OF HEAVY TAILS
insurance company. We conclude that the usual estimation procedure for the mean
value should not be used in the Pareto case. The same is true for other heavy tailed
distribution functions.
Let us next consider a reinsurance problem. A company wants to reinsure
Pa(,1) distributed claims via an excess of loss reinsurance with retention level
M. In order to x the reinsurance premium the mean value of the part of a claim
involving the reinsurer is needed.
II E[Y
i
M [ Y
i
> M] =
_
M
x
(1+x)
+1
dx
_
(1+x)
+1
dx
M =
_
(1+x)
dx
_
(1+x)
+1
dx
(M + 1)
=
1
(1 +M)
(1)
(1 +M)
(M + 1) =
(M + 1)
1
(M + 1)
=
M + 1
1
.
It follows that II E[Y
i
M [ Y
i
> M] increases linearly with M. Thus in order to
estimate II E[Y
i
M [ Y
i
> M] only the large values Y
i
> M should be used for the
estimation procedure. But this event might be rare. In contrast consider the Exp(1)
case. Here II E[Y
i
M [ Y
i
> M] = 1 for all M and thus the mean value needed can
be inferred from the full sample.
An argument often heard is the following: The aggregate value of all goods on
earth is very large, but nite. Thus a claim must be bounded which means light
tailed. Therefore there is no need to consider the heavy tailed case. But this
contradicts the actuarial experience. Quite often claim amounts show a behaviour
typical to heavy tails. For instance considering damages due to hurricanes shows
that the aggregate sum of claims is determined by a few of the largest claim. Thus
an insurance company has to be prepared that the next very large claim is much
larger. Moreover the aggregate value of all goods is so large that it can be considered
to be innite.
We conclude that one has to be careful as soon as heavy tailed claims are involved.
Actuaries use the phrase heavy tailed claims are dangerous. Thus one has to
treat the heavy tailed claim case specially.
6. SOME ASPECTS OF HEAVY TAILS 121
1
2
7
6
4
3
5
1 2 3 4 5
1
2
3
4
5
Figure 6.2: Mean residual life functions: (1) Exp(1), (2) (3,1), (3) (0.5,1), (4)
Wei(2,1), (5) Wei(0.7,1), (6) LN(-0.2,1), (7) Pa(1.5,1).
6.2. How to Find Out Whether a Distribution Is Heavy Tailed or
Not
6.2.1. The Mean Residual Life
We considered before the quantity II E[Y
i
M [ Y
i
> M], which is called mean
residual life in survival analysis. Let us consider some other distributions and the
limits for M . We rst note that
II E[Y
i
M [ Y
i
> M] =
_
M
_
y
M
dz dF(y)
1 F(M)
=
_
M
(1 F(z)) dz
1 F(M)
.
i) Exp()
II E[Y
i
M [ Y
i
> M] =
1
1
.
ii) (, )
II E[Y
i
M [ Y
i
> M] =
_
M
_
z
y
1
e
y
dy dz
_
M
y
1
e
y
dy
1
.
iii) Wei(, c)
II E[Y
i
M [ Y
i
> M] =
_
M
e
cz
dz
e
cM
_
0, if > 1,
c
1
, if = 1,
, if < 1.
122 6. SOME ASPECTS OF HEAVY TAILS
iv) LN(,
2
)
II E[Y
i
M [ Y
i
> M] =
_
M
_
log z
_
dz
_
log M
_ .
v) Pa(, )
II E[Y
i
M [ Y
i
> M] =
M +
1
.
Figure 6.2 shows the mean residual life functions of several distribution functions.
From the computations above we can see that the mean residual life tends to innity
as M in the heavy tailed cases, and to a nite value in the small claim cases.
One can also show that in the case of a subexponential distribution the mean residual
life function tends to innity as M .
The idea of the approach is to estimate the mean residual life function. If the
function is increasing to innity, then the distribution function should be consid-
ered to be heavy tailed. If it is converging to a nite constant then we can model
it to be light tailed.
6.2.2. How to Estimate the Mean Residual Life Function
Let Y
i
: 1 i n be a sample of iid. random variables. The natural estimator of
II E[Y
i
M [ Y
i
> M] is
e(M) =
1
n
i=1
1I
Y
i
>M
n
i=1
1I
Y
i
>M
(Y
i
M) .
e(M) is called empirical mean residual life function. We have seen before that
it is dangerous to estimate the mean value in this way, especially if the claim sizes
are heavy tailed. But fortunately, we are not really interested in the real mean
residual life function. We are only interested in the shape of its graph. And for this
purpose the estimator e(M) is good enough.
The following pictures show histograms and empirical residual life functions gen-
erated by a simulated sample of size 1000 each.
Exp(1)
1 2 3 4 5 6
7
50
100
150
1 2 3 4 5 6
0.6
0.7
0.8
0.9
6. SOME ASPECTS OF HEAVY TAILS 123
(3, 1)
2 4 6 8 10 12
10
20
30
40
50
60
2 4 6 8 10
1.5
2.0
2.5
3.0
(0.5, 1)
1 2 3 4 5
50
100
150
200
250
300 1 2 3 4 5
0.2
0.4
0.6
0.8
Wei(2,1)
0.5 1.0 1.5 2.0 2.5
20
40
60
80
0.5 1.0 1.5 2.0
0.2
0.4
0.6
0.8
Wei(0.7,1)
5 10 15 20
50
100
150
200
250
0 2 4 6 8 10 12
2
3
4
5
6
7
8
124 6. SOME ASPECTS OF HEAVY TAILS
LN(-0.2,1)
5 10 15 20
50
100
150
5 10 15
2
3
4
5
6
7
8
Pa(1.5,1)
20 40 60 80 100
100
200
300
400
500
600
10 20 30 40 50 60 70
10
15
20
25
30
35
One can see clearly that the exponential, the two gamma and the rst Weibull
distributions are light tailed. The lognormal and the Pareto distributions are clearly
heavy tailed (note that the last few points are not signicant because only few data
were used for its calculation). The second Weibull distribution yields problems.
Though the empirical mean residual life function is increasing, it looks like it would
converge. On the other hand there are points far out. The largest value is 20.6,
the largest but one is 12.4. This behaviour is typical for heavy tailed data. This
example shows that it is important also to consider a plot of the data in order to
nd the right model.
To end this section we consider real data from Swedish and Danish re insurance.
Example 6.1. (Swedish re insurance) The sample consists of 218 claims for
1982 in units of Millions of Swedish kroner.
5 10 15 20 25 30 35
20
40
60
80
5 10 15 20 25 30
4
6
8
10
12
6. SOME ASPECTS OF HEAVY TAILS 125
The histogram (some claims are far out) and the mean residual life function (which is
clearly increasing) indicate that the distribution function is heavy tailed. The empir-
ical mean residual life function is rst decreasing and after that linearly increasing.
This indicates that the larger claims follow a Pareto or a lognormal distribution.
Example 6.2. (Danish re insurance) The sample consists of 500 large claims
from 1st January 1980 till 31st December 1990. The unit is millions of Danish
Kroner (1985 prices). The data are also analysed in [68] and [36].
50 100 150 200 250
10
20
30
40
50
60
70
20 40 60 80 100 120 140
20
40
60
80
100
120
As in Example 6.1 the histogram and the mean residual life function indicate that
the distribution function is heavy tailed. Most likely the larger claims follow a Pareto
or a lognormal distribution.
6.3. Parameter Estimation
If a model for the data is chosen then it remains to estimate the parameters.
6.3.1. The Exponential Distribution
For the exponential distribution both the maximum likelihood estimator and the
method of moments yield the estimator
=
_
1
n
n
i=1
Y
i
_
1
.
6.3.2. The Gamma Distribution
The log likelihood function is
n
i=1
( log log () + ( 1) log Y
i
Y
i
) .
126 6. SOME ASPECTS OF HEAVY TAILS
This yields the equations
=
n
n
i=1
Y
i
and
log
t
()
()
+
1
n
n
i=1
log Y
i
= 0 .
These equations can only be solved numerically.
Because Gamma distributions are light tailed the method of moments can be
used. This yields the estimators
=
Y
1
n1
n
i=1
(Y
i
Y )
2
and
=
Y
2
1
n1
n
i=1
(Y
i
Y )
2
where
Y =
1
n
n
i=1
Y
i
. Because of the simplicity the second estimator should be
prefered.
6.3.3. The Weibull Distribution
The log likelihood function is
n
i=1
(log + log c + ( 1) log Y
i
cY
i
) .
We can assume that ,= 1. The we obtain the equations
c
1
n
n
i=1
Y
i
= 1
and
n
n
i=1
(cY
i
1) log Y
i
= 1 .
Also these equations have to be solved numerically.
The equations obtained from the method of moments are even harder to solve.
Anyway, taking into account that the distribution is heavy tailed for < 1, the
method of moments should not be used, except if it is clear that > 1.
6. SOME ASPECTS OF HEAVY TAILS 127
6.3.4. The Lognormal Distribution
Here obviously the best unbiased estimator is (compare with the normal distribution)
=
1
n
n
i=1
log Y
i
and
2
=
1
n 1
n
i=1
(log Y
i
)
2
.
6.3.5. The Pareto Distribution
Solving the maximum likelihood equations yields the two equations
n
n
i=1
log(1 +Y
i
/) = 1
and
1 =
+ 1
n
n
i=1
(1 +/Y
i
)
1
.
Also these equations have to be solved numerically.
We shall try to nd another approach. Assume for the moment that is known.
Then the maximum likelihood estimator for
=
_
1
n
n
i=1
log(1 +Y
i
/)
_
1
looks similar to the estimator in the exponential case. In fact,
II P[log(1 +Y
i
/) > x] = II P[Y
i
> (e
x
1)] = e
x
is exponentially distributed. Because, for Y
i
large, log(1 + Y
i
/) log(Y
i
/) =
log Y
i
log , we nd that intuitively log Y
i
is approximately Exp() distributed for
large values of Y
i
. Explicitly
II P[log Y
i
> x] = (1 + e
x
/)
= (e
x
+ 1)
e
x
e
x
and
II P[log Y
i
> M +x [ log Y
i
> M] =
_
e
Mx
+ 1
e
M
+ 1
_
e
x
.
128 6. SOME ASPECTS OF HEAVY TAILS
If we choose M large enough then
=
_
1
n
i=1
1I
log Y
i
>M
n
i=1
1I
log Y
i
>M
(log Y
i
M)
_
1
is an estimator for .
There remains one problem: How shall we choose M?
The larger M the closer is the distribution of log Y
i
M to an exponential
distribution.
The larger M the less data are used in the estimator.
It is clear that M can be chosen to be large if n is larger. Thus M must depend on
n. Moreover the optimal M will depend on the parameters and , which are not
known. Thus M must depend on the sample as well. This can be achieved if we
estimate only using the largest k(n) + 1 data. Let Y
i:n
: i n denote the order
statistics Y
1:n
Y
2:n
Y
n:n
of Y
i
: i n. We dene the estimator
=
_
1
k(n)
k(n)
i=1
(log Y
n+1i:n
log Y
nk(n):n
)
_
1
. (6.2)
Intuitively, this estimator will converge to as n provided k(n) and
Y
k(n):n
a.s.. It can be shown that whenever k(n)/n 0 and k(n)/ log log n
then a.s..
It remains to estimate . Note that for n and (0, 1)
1 (1 +Y
]n|:n
/)
a.s.,
which we will see later (Section 6.4.1). Thus for n large we nd the estimator
=
Y
](1)n|:n
1/
1
.
Because quite often the distribution will not be explicitly Pareto, only the tail will
behave like the tail of a Pareto distribution, we wish to let 0 as n . Choose
therefore n = k(n), i.e. our estimator is
=
Y
nk(n):n
(n/k(n))
1/
1
.
In order to simplify the estimator note that asymptotically this is the same as
=
_
k(n)
n
_
1/
Y
nk(n):n
. (6.3)
6. SOME ASPECTS OF HEAVY TAILS 129
20 40 60 80 100
1.2
1.4
1.6
1.8
2.0
Figure 6.3: Hills estimator as function of k for the Swedish re insurance data.
One can show that
a.s. as n .
The last question of interest is how to choose k(n) optimally. It can be shown
that in many situations the following holds. If, for any > 0, k(n)n
2/3+2
converges
to 0 then the rate of convergence of to is n
1/3+
. Thus, in practice, one would
take as small as possible and thus use k(n) = n
2/3
|. Another possibility is to
plot the estimator for k = 1, 2, 3 . . .. Then choose the rst point, where the function
(k) stabilises (for a short time). More about this and also other estimators can be
found in [36].
6.3.6. Parameter Estimation for the Fire Insurance Data
We now t a Pareto and a lognormal model to the Swedish and Danish re insurance
data considered before. We already have seen that a heavy tailed distribution func-
tion should be used in order to model the re insurance data. The explicit gures
for the claims have not been given here, so the reader cannot do the calculations
himself.
Example 6.1 (continued). In the Pareto case we choose k(218) = 218
2/3
| = 36.
Figure 6.3 indicates that k indeed should be chosen close to 36. We nd the estimates
= 1.37567,
= 0.879807.
Note that we nd the mean value
/( 1) = 2.34199, whereas the empirical mean
value is 2.28172, only slightly smaller. The big dierence lies in the estimate for
130 6. SOME ASPECTS OF HEAVY TAILS
50 100 150 200
1.4
1.5
1.6
1.7
1.8
1.9
Figure 6.4: Hills estimator as function of k for the Danish re insurance data.
the variance. The empirical variance is 14.3158, whereas in the Pareto model the
variance is innite.
Fitting a lognormal model we nd the parameters
= 0.218913,
2
= 0.945877.
In the lognormal model we obtain for the mean value 1.99741 which is even smaller
than the empirical mean value. The variance of the lognormal model is 6.28397,
which again is smaller than the empirical variance. This shows that choosing a
Pareto model the insurance company is on the safer side. Therefore the Pareto
model is very popular among practitioners. The comparison with the empirical
mean value and the empirical variance shows that we should not choose this model.
The problem we have with the lognormal model is that there are too many small
claims, i.e. a real lognormal distribution would have more claims close to e
. These
small claims are responsible that is relative small. The Pareto distribution has a lot
of very small claims, but for the estimation they are not considered. That is another
advantage of the Pareto model. For the estimation only the tail is considered. And
the tail is the important part of the distribution function in practice.
Example 6.2 (continued). In the Pareto case we choose k(500) = 500
2/3
| = 62.
Figure 6.4 shows that the Hill estimator as a function of k is very unstable. For a
short range it stabilizes around 70. Thus 62 seems not to be a too bad choice. It
starts really to stabilize at about k = 170. This seems to be too large for k. Anyway,
we calculate the parameters for both k = 62 and k = 180.
6. SOME ASPECTS OF HEAVY TAILS 131
For k = 62 the estimates are
= 1.69605,
= 4.20413.
The mean value is 6.03997, whereas the empirical mean was 9.08176. This indicates
that k = 62 is not a good choice. However, the Pareto distribution will have a lot
of very small claims, which here is not the case. We should here only model the
claims over a certain threshold to be Pareto distributed. For example, if we only
take claims over 10 into account, i.e. the largest 109 claims the empirical mean will
be 24.0818, whereas the chosen parametrization will yields a conditional mean of
30.4067.
For k = 180 the estimates are
= 1.34103,
= 2.87879.
The mean value is 8.4415, which still is smaller than the empirical mean value.
Taking only values over 10 into account the conditional mean will be 47.7646, which
is much larger than the empirical conditional mean 24.0818.
Fitting a lognormal distribution the parameters are
= 1.84616,
2
= 0.461025.
In the lognormal model we obtain for the mean value 7.97787 which is also much
smaller than the empirical mean value. The variance of the lognormal model is
100.924, which again is much smaller than the empirical variance 270.933. This
shows that choosing a Pareto model the insurance company is on the safer side.
6.4. Verication of the Chosen Model
6.4.1. Q-Q-Plots
Let Y
i
: 1 i n be iid. with distribution function G, and for simplicity we
assume that G is continuous. Consider the random variables G(Y
i
). It is clear
that the G(Y
i
)s are iid.. We want to nd the distribution function of G(Y
i
).
Let G
1
: (0, 1) IR, x infy IR : G(y) > x denote the generalised
inverse function of G. Let x (0, 1). Because G is increasing and continuous we
nd that G(G
1
(x)) = x and G(z) x z G
1
(x). Therefore
II P[G(Y
i
) x] = II P[Y
i
G
1
(x)] = G(G
1
(x)) = x .
132 6. SOME ASPECTS OF HEAVY TAILS
It follows that G(Y
i
) is uniformly distributed on (0, 1).
Let X
i
= G(Y
i
). Considering the order statistics it is easy to see that X
i:n
=
G(Y
i:n
). The distribution function of X
k:n
is
II P[X
k:n
> x] =
k1
i=0
II P[X
i:n
x, X
i+1:n
> x] =
k1
i=0
_
n
i
_
x
i
(1 x)
ni
.
From this we can compute the rst two moments of X
k:n
.
II E[X
k:n
] =
_
1
0
k1
i=0
_
n
i
_
x
i
(1 x)
ni
dx =
k1
i=0
_
n
i
__
1
0
x
i
(1 x)
ni
dx
=
k1
i=0
_
n
i
_
(i + 1)(n i + 1)
((i + 1) + (n i + 1))
=
k1
i=0
1
n + 1
=
k
n + 1
.
II E[X
2
k:n
] =
_
1
0
2x
k1
i=0
_
n
i
_
x
i
(1 x)
ni
dx = 2
k1
i=0
_
n
i
__
1
0
x
i+1
(1 x)
ni
dx ,
= 2
k1
i=0
_
n
i
_
(i + 2)(n i + 1)
(n + 3)
=
2
(n + 1)(n + 2)
k1
i=0
(i + 1)
=
k(k + 1)
(n + 1)(n + 2)
.
Thus the variance of X
k:n
is
Var[X
k:n
] =
k(n + 1 k)
(n + 1)
2
(n + 2)
1
4(n + 2)
.
The variance is therefore converging to 0 as n , i.e. X
k:n
lies close to k/(n+1).
The result can be made more precise. Because [0, 1] is compact and the pointwise
limit distribution is continuous, we have that sup[X
k:n
k/(n + 1)[ : k n
converges to zero.
Consider now the graph generated by the points (k/(n+1), G(Y
k:n
)). The points
are expected to lie close to the line y = x. This graph is called Q Q plot
(quantile-quantile-plot).
We now consider the Q-Q-plots of the simulated data verse the correct distribu-
tions.
6. SOME ASPECTS OF HEAVY TAILS 133
Exp(1)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
(0.5, 1)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
Wei(0.7,1)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
Pa(1.5,1)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
(3, 1)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
Wei(2,1)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
LN(-0.2,1)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
We see that the graphs are very close to the line y = x. Thus the random number
generator yields satisfactory results.
134 6. SOME ASPECTS OF HEAVY TAILS
If we did not choose the right distribution then the following may happen.
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
G(Y
k:n
) is too large. The correct function
G(x) should be smaller than the chosen one,
i.e. the chosen distribution has too much
weight close to 0. The correct distribution is
therefore less skewed than the chosen one.
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
G(Y
k:n
) is too small. The correct function
G(x) should be larger than the chosen one,
i.e. the chosen distribution has not enough
weight close to 0. The correct distribution
is therefore more skewed than the chosen
one.
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
G(Y
k:n
) is too large for small k and too small
for large k. The correct distribution should
have less small values and less very large val-
ues, i.e. the correct distribution should have
a lighter tail than the chosen distribution.
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
G(Y
k:n
) is too small for small k and too large
for large k. The correct distribution should
have more small values and more very large
values, i.e. the correct distribution should
have a thicker tail than the chosen distribu-
tion.
Remarks.
i) Note that the Q-Q-plot is sensitive to the parameter estimation. A bad Q-Q-plot
must not mean that the model is wrong, it can also mean that the parameters
6. SOME ASPECTS OF HEAVY TAILS 135
have been estimated badly.
ii) In order to make the right decision the sample size has to be taken into account.
iii) Note that the picture only depends on the sample and not on the whole distri-
bution. If for instance a class of distribution functions is chosen with a too light
tail, then the large values of the sample will force the estimated parameters to
be from a skewed model. The Q-Q-plot will then suggest that the tail should be
lighter than the chosen one even though it is vice versa.
6.4.2. The Fire Insurance Example
Let us now consider the re insurance example again. We had tted a Pareto and
a lognormal model to the data.
Example 6.1 (continued). The Q-Q-plots in this example are
Pa(1.37567, 0.879807)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
LN(0.218913, 0.945877)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
All investigations so far had indicated that the underlying distribution is Pareto.
But now the Q-Q-plot suggests that the lognormal distribution is closer to the true
distribution. The Pareto distribution is too much skewed. The reason might be
that the -parameter is too far away from the true value. On the other hand the
lognormal distribution seems to t quite well in the tail. It is not too far away from
the right distribution for smaller values. We can of course not expect that one of
the considered distributions ts perfectly. But we try to nd a distribution with
a tail similar to the correct distribution function. The Q-Q-plot suggest that the
lognormal distribution has the right tail.
We conclude that it is important to consider a larger set of distribution functions,
even if some evidence shows that another distribution function should be prefered.
As in our case it might turn out that not the preferred distribution function is the
136 6. SOME ASPECTS OF HEAVY TAILS
correct one. One of the problems here would be that 218 is not a very large sample
size. In practice one would be able to collect data for several years, which might
help to get a clearer picture of the distribution.
Example 6.2 (continued).
The Q-Q-plots in this example are
Pa(1.69605, 4.20413)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
LN(1.84616, 0.461025)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
Pa(1.34103, 2.87879)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
We clearly see that there are not enough very small data, but far out in the tails
of the Pareto distributions seem to be right. The lognormal distribution clearly is
not the right distribution.
Because the absence of the very small claims cause problems we should have
considered a conditional Pareto distribution instead. Condition on Y
i
> Y
1:n
. Then
the Q-Q-plots are
6. SOME ASPECTS OF HEAVY TAILS 137
Pa(1.69605, 4.20413)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
Pa(1.34103, 2.87879)
0.2 0.4 0.6 0.8 1.0
0.2
0.4
0.6
0.8
1.0
The chosen parametrization both do not have enough weight close to 0 and near
innity. This indicates that the correct parameter should be chosen smaller than
the estimated ones or the correct parameter is larger than the estimated one. This
was also indicated by the fact that the mean value of the estimated distributions
were smaller than the empirical mean value.
Bibliographical Remarks
The graphical tool for recognizing heavy tailed distributions, the mean residual life
function, can be found for instance in [54] or [37]. The Swedish re insurance data
are taken from [57]. The data can also be found in [37]. The estimator (6.2) was
introduced by Hill [53] and is therefore often called the Hill estimator. That the
estimators (6.2) and (6.3) are strongly consistent was proved in [29]. The optimal
choice of k(n) can be found in [52]. A comprehensive introduction into the eld can
be found in the book [36]. This book mainly discusses the approach via extreme
value theory.
138 7. THE AMMETER RISK MODEL
7. The Ammeter Risk Model
7.1. Mixed Poisson Risk Processes
We have seen that a continuous time risk model is easy to handle if the claim
number process is a Poisson process. Considering a risk in a xed time interval we
learned that a compound Poisson model does not t real data very well. We had
to use a compound mixed Poisson model, as for instance the compound negative
binomial model. The reason that these models are closer to reality is that they allow
more uctuation.
It is natural to ask whether we can allow for more uctuation in the Poisson
process. We already saw one way to allow for more uctuation, that is the renewal
risk model. But this is not an analogue of the compound mixed Poisson model in a
xed time interval. The easiest way in order to obtain a mixed Poisson distribution
for the number of claims in any nite time interval is to let the claim arrival intensity
to be stochastic. Such a model is called mixed Poisson risk process.
Let H(x) be the distribution function of . Then, using the usual notation,
(u) =
_
0
c
(u, ) dH() =
_
c/
0
c
(u, ) dH() + (1 H(c/))
where
c
(u, ) denotes the ruin probability in the Cramer-Lundberg model with
claim arrival intensity . It follows immediately that there can only exist an ex-
ponential upper bound if there exists an
0
such that H(
0
) = 1 and c >
0
.
Otherwise, as for instance if is gamma distributed, Lundbergs inequality or the
Cramer-Lundberg approximation cannot hold. In particular, there is no possibility
to formulate a net prot condition without knowing an upper bound for .
Assume we have observed the process up to time t. We want to show that
II P[ [ T
t
] = II P[ [ N
t
]. Because N
t
and Y
i
are independent it follows
that II P[ [ T
t
] = II P[ [ N
t
, T
1
, T
2
, . . . , T
Nt
]. Recall that, given N
t
= n
the claim times T
1
, T
2
, . . . , T
n
have the same distribution as the order statistics of n
independent uniformly in [0, t] distributed random variables (Proposition C.2). Let
7. THE AMMETER RISK MODEL 139
n IIN 0 and let 0 t
i
t for 1 i n. Then
II P[ , N
t
= n, T
1
t
1
, . . . , T
n
t
n
]
= II P[ , T
1
t
1
, . . . , T
n
t
n
, T
n+1
> t]
=
_
0
_
t
1
0
xe
xs
1
_
t
2
s
1
s
1
xe
x(s
2
s
1
)
_
tns
n1
s
n1
xe
x(sns
n1
)
e
x(tsn)
ds
n
ds
1
dH(x)
=
_
0
(xt)
n
n!
e
xt
_
t
1
0
_
t
2
s
1
s
1
_
tns
n1
s
n1
n!t
n
ds
n
. . . ds
1
dH(x)
= II P[ , N
t
= n]II P[T
1
t
1
, . . . , T
n
t
n
[ N
t
= n]
= II P[ [ N
t
= n]II P[T
1
t
1
, . . . , T
n
t
n
, N
t
= n]
and it follows that
II P[ [ N
t
= n, T
1
t
1
, . . . , T
n
t
n
] = II P[ [ N
t
= n] .
Thus II P[ [ T
t
] = II P[ [ N
t
].
We know from the renewal theory (using the strong law of large numbers) that,
given , t
1
N
t
converges to a.s.. Thus
II P[ lim
t
t
1
N
t
= ] = 1 .
Moreover, Var[ [ N
t
] 0 as t . Thus as time increases the desired variability
disappears. Hence this model is not the desired generalisation of the mixed Poisson
model in a nite time interval.
7.2. Denition of the Ammeter Risk Model
In 1948 the Swiss actuary Hans Ammeter had the idea to combine a sequence of
mixed Poisson models. He wanted to start a new independent model at the beginning
of each year. That means a new intensity level is chosen at the beginning of each
year independently of the past intensity levels. This has the advantage that the
variability does not disappear with time.
For a formal denition let denote the length of the time interval in which the
intensity level remains constant, for instance one year. Let L
i
be a sequence of
positive random variables with distribution function H. We allow for the possibility
that H(0) ,= 0. The intensity at time t is
t
:= L
i
, if (i 1) t < i.
140 7. THE AMMETER RISK MODEL
The stochastic process
t
is called claim arrival intensity process of the risk
model. Dene
t
=
_
t
0
s
ds and let
N
t
be a homogeneous Poisson process with
rate 1. The claim number process is then
N
t
:=
N
t
.
Intuitively the point process N
t
can be obtained in the following way. First
generate a realization
t
. Given
t
: 0 t < let N
t
be an inhomogeneous
Poisson process with rate
t
(compare with Proposition C.5). A process obtained
in this way is called doubly stochastic point process or Cox process. The
process N
t
constructed above and the mixed Poisson process are special cases of
Cox processes.
As for the models considered before let the claims Y
i
be a sequence of iid. pos-
itive random variables with distribution function G. The Ammeter risk model is the
the process
C
t
= u +ct
Nt
i=1
Y
i
where u is the initial capital and c is the premium rate. As before let := inft >
0 : C
t
< 0 denote the time of ruin and (u) = II P[ < [ C
0
= u] denote the ruin
probability.
In order to investigate the model we dene also the probability of ruin obtained
by considering only the times , 2, . . . where the intensity changes, i.e. let
=
infi : i IIN, C
i
< 0 and
(u) := II P[
t
= u +ct
N
i=1
Y
i
where N
t
is a renewal process with interarrival time distribution F
(t) = 1I
t
and claim sizes Y
i
with a compound mixed Poisson distribution. More explicitly,
G
(x) = II P[Y
i
x] = II P
_
N
i=1
Y
i
x
_
.
Given L
1
the conditional expected number of claims and thus the Poisson parameter
is L
1
. From (1.2) we already know the moment generating function
M
Y
(r) = M
L
((M
Y
(r) 1))
7. THE AMMETER RISK MODEL 141
of Y
i
where M
L
(r) is the moment generating function of the intensity levels L
i
.
In order to avoid (u) = 1 we need that the random walk C
k
tends to innity,
or equivalently that II E[C
t
can be compared in the following
way.
Lemma 7.1. For (u) and
(u) (u)
(u c) .
Proof. Assume that there exists t
0
> 0 such that C
t
0
+ c < 0, i.e. ruin occurs
for the Ammeter risk model with initial capital u + c. Let k
0
such that t
0
((k
0
1), k
0
]. Then
0 > c +C
t
0
= c +C
k
0
c(k
0
t
0
) +
N
k
0
i=Nt
0
+1
Y
i
> C
k
0
.
Thus (u + c)
= C
k
0
k
.
These points are exactly the claim times of the renewal risk process C
k
. The
assertion follows from Lemma 5.1 and from (5.3).
Because (r) is a convex function and (0) = 0 there may exist a second solution
R to the equation (r) = 0. R is, if it exists, unique and strictly positive. R is then
called the Lundberg exponent or adjustment coecient of the Ammeter risk
model.
Next we proof Lundbergs inequality.
Theorem 7.3. Let C
t
be an Ammeter risk model and assume that the Lundberg
exponent R exists. Then
(u) < e
cR
e
Ru
.
If there exists an r > R such that M
L
((M
Y
(r) 1)) < then R is the right
exponent in the sense that there exists a constant C > 0 such that
(u) Ce
Ru
.
Remark. The constant e
cR
might be too large, especially if is large. If
possible, one should therefore try to nd an alternative for the constant. An upper
bound, based on the martingale (7.2), is found in [35] and [66]. However, one has to
accept that the constant might be larger than 1. Recall that this was also the case
for Lundbergs inequality in the general case of the renewal risk model.
Proof. Using Lemma 7.1 the proof becomes almost trivial. From Theorem 5.5 it
follows that
(u) < e
Ru
. Thus
(u)
(u c) < e
R(uc)
= e
cR
e
Ru
.
7. THE AMMETER RISK MODEL 143
If there exist an r > R such that M
L
((M
Y
(r) 1)) < then by Theorem 5.7
there exists a constant C > 0 such that
(u) Ce
Ru
. Thus also
(u)
(u) Ce
Ru
.
The constant e
cR
seems to increase to innity as tends to innity. But R
also dependends on . At least in some cases e
cR
remains bounded as .
Proposition 7.4. Assume that R exists and that there exists a solution r
0
,= 0 to
M
L
(r)e
cr
= 1 .
Then
e
cR
e
cr
0
. (7.1)
In particular, the condition will be fullled if R exists and if there exists
such that M
L
(r) < r <
.
Remark. Note that r
0
depends on the claim size distribution via only.
Proof. The function log(M
L
(r)) cr has second derivative
2
_
M
tt
L
(r)
M
L
(r)
_
M
t
L
(r)
M
L
(r)
_
2
_
> 0
(see Lemma 1.9) and is therefore strictly convex. Because
_
M
t
L
(0)
M
L
(0)
_
c = II E[L
i
] c < 0
r
0
has to be strictly positive. In particular, if
M
L
(r)e
cr
= explog(M
L
(r)) cr 1
then r r
0
. If M
L
(r) < r <
) = .
Thus by the convexity a strictly positive solution to log(M
L
(r)) cr = 0 has to
exist.
Recall that M
L
(r) is increasing and that M
Y
(r) 1 rM
t
Y
(0) = r. Since
1 = e
cR
M
L
((M
Y
(R) 1)) e
cR
M
L
(R)
it follows from the consideration above that R r
0
.
144 7. THE AMMETER RISK MODEL
Example 7.5. Assume that L
i
(, ). The case = 1 was the case considered
in Ammeters original paper [2]. In order to obtain the Lundberg exponent we have
to solve the equation
e
cr
_
(M
Y
(r) 1)
_
= 1
or equivalently
e
cr/
(M
Y
(r) 1)
= 1 .
This can be written as
[ (M
Y
(r) 1)] e
cr/
= .
It follows immediately that the above equation admits only a solution in closed form
for special cases. For example, if Y
i
= 1 is deterministic then one has to solve
e
(1+c/)r
( + )e
(c/)r
+ = 0 .
One can only hope for a solution in closed form if c/ Q. For instance if
c/ = 1 then R = log(/) = log(c/).
Let us now consider (7.1). r
0
is the strictly positive solution to
_
r
_
e
cr
= 1
or
( r)e
cr/
= .
The latter equation also has to be solved numerically.
From Theorem 7.3 we nd that
C lim
u
(u)e
Ru
lim
u
(u)e
Ru
e
cR
.
The question is: exists there a constant C such that
lim
u
(u)e
Ru
= C ,
i.e. does the Cramer-Lundberg approximation hold? This question is answered in
the following theorem.
7. THE AMMETER RISK MODEL 145
Theorem 7.6. Let C
t
be an Ammeter risk model. Assume that R exists and
that there exists an r > R such that M
L
((M
Y
(r) 1)) < . Then there exists a
constant 0 < C e
cR
such that
lim
u
(u)e
Ru
= C .
Proof. This is a special case of Theorem 2 of [71].
We next want to derive an alternative proof of Lundbergs inequality. For this
reason we have to nd an exponential martingale in continuous time. In applications
it is usually rather easy to nd a martingale if one deals with Markov processes. But
we have remarked before that C
t
is not a Markov process. The future of the process
is dependent on the present level
t
of the intensity. Let us therefore consider the
process (C
t
,
t
). But this is not a homogeneous Markov process neither. The
future of the process is also dependent on time because the future intensity level
is known until the next change of the intensity. Instead of considering (C
t
,
t
, t)
we consider (C
t
,
t
, V
t
) where V
t
is the time remaining till the next change of the
intensity, i.e.
V
t
= kt, if (k 1) t < k.
Remarks.
i) The process
t
cannot be observed directly. The only information on
t
are
the number of claims occured in the interval [t +V
t
, t]. Thus, provided L
i
is
not deterministic, we can, in the best case, nd the distribution of
t
given T
C
t
,
the information generated by C
s
: s t up to time t. But for our purpose
it is no problem to work with the ltration T
t
generated by (C
t
,
t
, V
t
). The
results will not depend on the chosen ltration.
ii) As an alternative we could consider the time elapsed since the last change of
the intensity W
t
= V
t
. For the present choice of the Markovization see also
Section 8.2.
We nd the following martingale.
Lemma 7.7. Let r such that M
L
((M
Y
(r) 1)) < . Then the process
M
r
t
= e
rCt
e
(t(M
Y
(r)1)cr(r))Vt
e
(r)t
(7.2)
is a martingale with respect to the natural ltration of (C
t
,
t
, V
t
).
146 7. THE AMMETER RISK MODEL
Proof. Let rst (k 1) s < t < k. Then
t
=
s
and
II E[M
r
t
[ T
s
] = II E
_
e
rCt
e
(t(M
Y
(r)1)cr(r))Vt
e
(r)t
T
s
= M
r
s
II E
_
e
r(CtCs)
T
s
e
(t(M
Y
(r)1)cr)(ts)
= M
r
s
e
(t(M
Y
(r)1)cr)(ts)
e
(t(M
Y
(r)1)cr)(ts)
= M
r
s
.
Let now (k 1) s < t = k. Then, using the denition of (r),
II E[M
r
t
[ T
s
] = M
r
s
II E
_
e
r(CtCs)
T
s
II E
_
e
(L
k+1
(M
Y
(r)1)cr(r))
e
(s(M
Y
(r)1)cr(r))(ts)
e
(r)(ts)
= M
r
s
e
(s(M
Y
(r)1)cr)(ts)
M
L
((M
Y
(r) 1))e
(cr+(r))
e
(s(M
Y
(r)1)cr)(ts)
= M
r
s
.
The assertion follows by iteration.
From the martingale stopping theorem we nd
e
Ru
= II E[M
R
0
] = II E[M
R
t
] II E[M
R
; t] .
By monotone convergence it follows that
e
Ru
II E[M
R
; < ] = II E[M
R
[ < ]
follows. Note that C
1I
0v
+ 1I
v>
_
H() .
We get the following version of Lundbergs inequality.
Theorem 7.8. Let C
t
be an Ammeter risk model. Assume that the Lundberg
exponent R exists and that
II E[e
0
(M
Y
(R)1)V
0
] < .
Then
(u) < e
cR
II E[e
(
0
(M
Y
(R)1)cR)V
0
]e
Ru
.
Remark. The condition
II E
_
e
0
(M
Y
(R)1)V
0
<
will be fullled in all reasonable situations. For instance, if (
0
, V
0
) is distributed
such that
t
becomes stationary, then
II E[e
0
(M
Y
(R)1)V
0
] =
1
_
0
M
L
(v(M
Y
(R) 1)) dv < M
L
((M
Y
(R) 1)) < .
= II E[M
R
0
] = II E[M
R
t
] II E[M
R
; t] .
The assertion follows now as before.
For the Cramer-Lundberg approximation we need the following
148 7. THE AMMETER RISK MODEL
Lemma 7.9. Assume that an r > R exists such that M
L
((M
Y
(r) 1)) < and
that
II E[e
0
(M
Y
(r)1)V
0
] < .
Then
lim
u
II P[ V
0
[ C
0
= u]e
Ru
= 0 .
Proof. Again using the stopping theorem we get
e
ru
II E[e
(
0
(M
Y
(r)1)cr(r))V
0
] = II E[M
r
0
] = II E[M
r
V
0
] II E[M
r
; V
0
] .
As before it follows that
II P[ V
0
] e
(rc+(r))
II E[e
(
0
(M
Y
(r)1)cr(r))V
0
]e
ru
< e
(rc+(r))
II E[e
0
(M
Y
(r)1)V
0
]e
ru
where we used that (r) > 0. Multiplying with e
Ru
and then letting u yields
the assertion.
The Cramer-Lundberg approximation in the general case can now be proved.
Theorem 7.10. Let C
t
be an Ammeter risk model. Assume that R exists and
that there exists an r > R such that both M
L
((M
Y
(r) 1)) < and
II E[e
0
(M
Y
(r)1)V
0
] < .
Then
lim
u
(u)e
Ru
= II E[e
(
0
(M
Y
(R)1)cR)V
0
] C
where C is the constant obtained in Theorem 7.6.
Proof. Let
o
(u) denote the ruin probability in the ordinary case and dene
B(x; , v) := II P[C
v
u x [
0
= , V
0
= v] .
Then
(u) =
_
=0
_
v=0
_
cv
y=u
o
(u +y)II P[ > v [
0
= , V
0
= v, C
v
= u +y]
B(dy; , v)F
1
(d, dv) + II P[ V
0
] .
7. THE AMMETER RISK MODEL 149
We multiply the above equation by e
Ru
. The last term tends to 0 by Lemma 7.9.
The rst term on the right hand side can be written as
_
=0
_
v=0
_
cv
y=u
o
(u +y)e
R(u+y)
II P[ > v [
0
= , V
0
= v, C
v
= u +y]
e
Ry
B(dy; , v)F
1
(d, dv) .
Noting that by Theorem 7.6
o
(u + y)e
R(u+y)
is uniformly bounded by a constant
C we see that
_
=0
_
v=0
_
cv
y=u
o
(u +y)e
R(u+y)
II P[ > v [
0
= , V
0
= v, C
v
= u +y]
e
Ry
B(dy; , v)F
1
(d, dv)
CII E[e
R(C
V
0
u)
] =
CII E[e
(
0
(M
Y
(R)1)cR)V
0
] < .
We can therefore interchange integration and limit. Because
o
(u +y)e
R(u+y)
tends
to C and II P[ > v [
0
= , V
0
= v, C
v
= u + y] tends to 1 as u the desired
limit is
lim
u
(u)e
Ru
= C
_
=0
_
v=0
_
cv
y=
e
Ry
B(dy; , v)F
1
(d, dv)
= II E[e
(
0
(M
Y
(R)1)cR)V
0
] C .
i
has a subexponential distribution. But it is also possible that
L
i
has a distribution which makes Y
i
subexponential. In any case we have the
following
Proposition 7.11. Assume that
1
II E[L
i
]
_
x
0
(1 G
(y)) dy
is subexponential. Then
lim
u
(u)
_
u
(1 G
(y)) dy
=
1
cII E[L
i
]
.
150 7. THE AMMETER RISK MODEL
Proof. From Theorem 5.12 the asymptotic behaviour in the assertion is true for
(u) replaced by
(u) tends to
1 as u . We have from Lemma 7.1
1
(u)
(u)
(u c)
(u)
.
Using Theorem 5.12 and Lemma F.2 we obtain
lim
u
(u c)
(u)
= lim
u
_
uc
(1 G
(y)) dy
_
u
(1 G
(y)) dy
= 1 .
This proves the assertion.
The next result deals with the tail of compound distributions with a regularly
varying tail.
Lemma 7.12. Let N be a random variable with values in IIN such that II E[N] < ,
Y
i
be a sequence of positive iid. random variables independent of N with distribu-
tion function G(y) and mean , and let
G(y) be the distribution function of
N
i=1
Y
i
.
Let > 1 and L(x) be a slowly varying function. Assume that
lim
x
x
(1 G(x))
L(x)
= , lim
n
n
II P[N > n]
L(n)
=
where and are positive constants. Then
lim
x
x
(1
G(x))
L(x)
=
+ II E[N].
Proof. See [79].
Next we link the distribution function H() of L
i
to the distribution function of
N
.
Lemma 7.13. Assume that there exists a > 0 and a slowly varying function
L(x) such that
lim
(1 H())
L()
= 0 .
Then
lim
n
n
II P[N
> n]
L(n)
= 0 .
7. THE AMMETER RISK MODEL 151
Proof. First note that for n 1
II P[N
> n] =
_
0
m=n+1
()
m
m!
e
dH()
=
_
0
m=n+1
_
0
_
(y)
m1
(m1)!
(y)
m
m!
_
e
y
dy dH()
=
_
0
_
y
m=n+1
_
(y)
m1
(m1)!
(y)
m
m!
_
e
y
dH() dy
=
_
0
(y)
n
n!
e
y
(1 H(y)) dy =
_
0
y
n
n!
e
y
(1 H(y/)) dy .
Choose now a e
1
. Then
_
an
0
y
n
n!
e
y
(1 H(y/)) dy
_
an
0
y
n
n!
e
y
dy
(an)
n+1
n!
e
an
n
n+1
n!
e
ann
because y
n
e
y
is increasing in the interval (0, n). By Stirlings formula there exists
K > 0 such that n! Kn
n+
1
2
e
n
from which it follows that
n
L(n)
1
_
an
0
y
n
n!
e
y
(1 H(y/)) dy
converges to 0 as n . The latter follows because x
L(n)
1
_
an
y
n
n!
e
y
(1 H(y/)) dy
converges to 0 because
n
L(n)
1
(1 H(an/)) =
_
a
_
(an/)
(1 H(an/))
L(an/)
L(an/)
L(n)
converges to 0.
152 7. THE AMMETER RISK MODEL
Lemma 7.14. Let L(x) be a locally bounded slowly varying function and let > 0.
Then
lim
y
_
0
x
y
e
x
L(x) dx
(y + 1)L(y)
= 1 .
Proof. There exists a constant K such that L(x) Kx
. Assume that y is so
large that L(y) > y
. Then
_
e
1
y
0
x
y
e
x
L(x)
(y + 1)
dx
_
L(y)
Ky
(y + 1)
_
e
1
y
0
x
y
e
x
dx <
Ky
y++1
e
(1+e
1
)y
(y + 1)
which converges to 0 as y by Stirlings formula. The last inequality follows
because x
y
e
x
is increasing on [0, y].
Analogously
_
2y
x
y
e
x
L(x)
(y + 1)
dx
_
L(y)
Ky
(y + 1)
_
2y
x
y
e
x/2
e
x/2
dx
Ky
(2y)
y
(y + 1)
e
y
2e
y
which again converges to 0 as y by Stirlings formula.
Because L(xy)/L(y) converges uniformly as y for e
1
x 2 it follows
that
lim
y
_
2y
e
1
y
x
y
e
x
L(x) dx
(y + 1)L(y)
= lim
y
_
2y
e
1
y
x
y
e
x
dx
(y + 1)
= 1
where the last equality follows from the considerations above with L(x) = 1.
Lemma 7.15. Assume that there exists a > 0 and a locally bounded slowly
varying function L(x) such that
lim
(1 H())
L()
= 1 .
Then
lim
n
n
II P[N
> n]
L(n)
=
.
Proof. Let us consider
II P[N
> n]
n
n!
e
L(/) d
n
n!
e
L(/) d
0
n
e
L(/)
_
(/)
(1H(/))
L(/)
1
_
d
_
0
n
e
L(/) d
.
7. THE AMMETER RISK MODEL 153
Choose x
0
such that
1
2
<
(/)
(1 H(/))
L(/)
< 1 +
2
for x
0
. Then
x
0
n
e
L(/)
_
(/)
(1H(/))
L(/)
1
_
d
_
0
n
e
L(/) d
<
2
.
The remaining term can be estimated denoting
L = maxL(x/) : x x
0
_
x
0
0
n
e
(1 H(/)) d +
_
x
0
0
n
e
L(/) d
_
0
n
e
L(/) d
x
n+1
0
+x
n+1
0
L
_
2x
0
n
e
L(/) d
_
1
2
_
n
x
+1
0
+x
0
L
_
2x
0
e
L(/) d
Choosing n large enough the latter can be made smaller than /2. Thus
lim
n
II P[N
> n]
_
n
n!
e
L(/) d
=
.
It follows from Stirlings formula that
lim
n
n
(n + 1)
n!
= 1
and the assertion follows from Lemma 7.14 noting that L(x/) is also a slowly
varying function.
We can now prove the asymptotic behaviour in the case where the heavier tail
of G and H is regularly varying.
Theorem 7.16. Assume that there exist a slowly varying function L(x) and pos-
itive constants , and with > 1 and + > 0 such that
lim
(1 H())
L()
= , lim
x
x
(1 G(x))
L(x)
= .
Then
lim
u
(u)u
1
L(u)
=
II E[L
i
] +()
(c II E[L
i
])( 1)
.
Proof. By Lemma 7.13 and Lemma 7.15
lim
n
n
II P[N
> n]
L(n)
=
.
From Lemma 7.12 one obtains
lim
x
x
(1 G
(x))
L(x)
= II E[L
i
] +
.
The result follows now from Proposition 7.11 and Karamatas theorem.
154 7. THE AMMETER RISK MODEL
Example 7.5 (continued). Assume that the claims are Pa(, ) distributed with
> 1. Then
lim
x
x
(1 G(x)) =
and
lim
x
x
(1 H(x)) = lim
x
x
1
e
x
()x
(+1)
= 0 .
Thus
lim
u
u
1
(u) =
/
(c //( 1))( 1)
=
(c( 1) )
.
(1 G(x)) = 0 .
Then it follows that
lim
u
u
1
(u) =
()
(c /( 1) )( 1)
=
()
(c( 1) )
.
= supr 0 : M
L
((M
Y
(r) 1)) < and assume that r
> R. We
have discussed R(y, y) in the classical case (Section 4.14). Let r
y
be the argument
maximizing r (r)y. We obtain the critical value
y
0
= (
t
(R))
1
=
_
M
t
L
((M
Y
(R) 1))M
t
Y
(R)
M
L
((M
Y
(R) 1))
c
_
1
and the two nite time Lundberg inequalities
II P[0 < yu [ C
0
= u] < e
(cry+(ry))
II E[e
(
0
(M
Y
(ry)1)cry(ry))V
0
]e
R(0,y)u
where R(0, y) > R if y < y
0
, and
II P[yu < < [ C
0
= u] < e
(cry+(ry))
II E[e
(
0
(M
Y
(ry)1)cry(ry))V
0
]e
R(y,)u
where R(y, ) > R if y > y
0
. The next result can be proved in the same way as
Theorem 4.24 was proved.
Theorem 7.18. Assume that R < r
. Then
u
y
0
in probability on the set < .
Bibliographical Remarks
For a comprehensive introduction to mixed Poisson processes see [49]. The Ammeter
risk process was introduced by Ammeter [2]. The process, as dened here, was
rst considered by Bjork and Grandell [18], where also Lundbergs inequality was
proved. The approach used here can be found in [48]. In particular, Lemma 7.1,
Proposition 7.4, Proposition 7.11, Lemma 7.13 and Theorem 7.16 are taken from
[48]. Lemma 7.7 was rst obtained in [35]. Lemma 7.15 goes back to [49]. The nite
time Lundberg inequalities were rst proved in [35].
156 8. CHANGE OF MEASURE TECHNIQUES
8. Change of Measure Techniques
In the risk models considered so far we constructed exponential martingales. Using
the stopping theorem we were able to prove an upper bound for the ruin probabilities.
But the technique did not allow to prove a lower bound or the Cramer-Lundberg
approximation. In the Ammeter risk model it had been possible to nd a lower
bound by means of a comparison with a renewal risk model. The reason that this
was possible was that the lengths of the intervals, in which the intensity was constant,
was deterministic. This made the interarrival times and the claim amounts in the
corresponding renewal risk model independent.
In order to be able to treat models where the intervals with constant intensity
are stochastic we have to nd another method. A possibility is always the following.
Let M
t
be a strictly positive martingale. Suppose that M
t
0 on = and
that M
t
is bounded on > t. Then
M
0
= II E[M
t
] = II E[M
; t] + II E[M
t
; > t] .
Now monotone convergence yields II E[M
; t] II E[M
; < ] = II E[M
[ < ]
.
The problem is, that in order to prove a Cramer-Lundberg approximation one needs
that M
and < are independent. This is ususally not an easy task to verify.
The method we will use here are change of measure techniques. We will see
that the exponential martingales can be used to change the measure. Under the
new measure the risk model will be of the same type, except that the net prot
condition will not be fullled and ruin will always occur within nite time. Thus
the ruin problem will become easier to handle.
We found the asymptotic behaviour of the ruin probability. But in the renewal
as well as in the Ammeter case we were not able to give an explicit formula for the
constant involved. And the results only are valid for ruin in innite time. If we are
interested in an exact gure for the (nite or innite time) ruin probability or for
the distribution of the severity of ruin then the only method available is simulation.
8. CHANGE OF MEASURE TECHNIQUES 157
But the event of interest is rare. One will have to simulate a lot of paths of the
process which will never lead to ruin. If dealing with the event of ruin in innite
time, then we will have to decide at which point we declare the path as not leading
to ruin because we have to stop the simulation within nite time.
The change of measure technique will give us the opportunity to express the
quantity of interest as a quantity under the new measure. But because under the
new measure ruin occurs almost surely the problems discussed above vanish. Thus
the change of measure technique provides a clever way to simulate.
8.1. The Cramer-Lundberg Case
The easiest model in order to illustrate the change of measure technique is the
classical risk model. Let C
t
be a Cramer-Lundberg process and assume that the
Lundberg exponent R exists and that there is an r > R such that M
Y
(r) < . We
have seen that for any r such that M
Y
(r) < the process
L
r
t
:= e
r(Ctu)
e
(r)t
is a martingale where (r) = (M
Y
(r) 1) cr. This martingale has the following
properties
L
r
t
> 0 for all t 0,
II E[L
r
t
] = II E[L
r
0
] = 1 for all t 0.
These properties are exactly the properties needed to dene an equivalent measure
on the -algebra T
t
. If A T
t
then
Q
r
[A] := II E
II P
[L
r
t
; A]
is a measure on T
t
equivalent to II P. In fact it looks as though the measure also
depends on t, but the next lemma shows that this is not the case. We suppose here
that T
t
is the smallest right continuous ltration such that C
t
is adapted. From
iii) of the lemma below we can see that it is possible to extend the measures Q
r
on
T
t
to the whole -algebra T.
Lemma 8.1.
i) Let t > s and A T
s
T
t
. Then II E
II P
[L
r
t
; A] = II E
II P
[L
r
s
; A].
158 8. CHANGE OF MEASURE TECHNIQUES
ii) Let T be a stopping time and A T
T
such that A T < . Then
Q
r
[A] = II E
II P
[L
r
T
; A] .
iii) C
t
is a Cramer-Lundberg process under the measure Q
r
with claim size distri-
bution
G(x) =
_
x
0
e
ry
dG(y)/M
Y
(r) and claim arrival intensity
= M
Y
(r). In
particular II E
Qr
[C
1
u] =
t
(r).
iv) If r ,= 0 then Q
r
and P are singular on the whole -algebra T.
Proof. i) This is an easy consequence of the martingale property
II E
II P
[L
r
t
; A] = II E
II P
[II E
II P
[L
r
t
; A [ T
s
]] = II E
II P
[II E
II P
[L
r
t
[ T
s
]; A] = II E
II P
[L
r
s
; A] .
ii) By iii) and Kolmogorovs extension theorem Q
r
can be extended uniquely to a
measure on T. We have by the stopping theorem and the monotone convergence
theorem
Q
r
[A] = lim
t
Q
r
[A T t] = lim
t
II E
II P
[L
r
t
; A T t]
= lim
t
II E
II P
[II E
II P
[L
r
t
[ T
T
]; A T t] = lim
t
II E
II P
[L
r
T
; A T t]
= II E
II P
[L
r
T
; A] .
iii) For all n 1 we have to nd the joint distribution of T
1
, T
2
T
1
, . . . , T
n
T
n1
and Y
1
, Y
2
, . . . , Y
n
. Note that these random variables are independent under II P. Let
t 0, A be a Borel set of IR
n
+
and B
1
, B
2
, . . . , B
n
be Borel sets on IR
+
. Then
Q
r
[N
t
= n, (T
1
, T
2
, . . . , T
n
) A, Y
1
B
1
, Y
2
B
2
, . . . , Y
n
B
n
]
= II E
II P
[L
r
t
; N
t
= n, (T
1
, T
2
, . . . , T
n
) A, Y
1
B
1
, Y
2
B
2
, . . . , Y
n
B
n
]
= II E
II P
_
exp
_
r
n
i=1
Y
i
_
exp((r) +cr)t; N
t
= n, (T
1
, . . . , T
n
) A,
Y
1
B
1
, Y
2
B
2
, . . . , Y
n
B
n
_
=
(t)
n
n!
e
t((r)+cr)t
II P[(T
1
, T
2
, . . . , T
n
) A [ N
t
= n]
n
i=1
II E
II P
[e
rY
i
; Y
i
B
i
]
=
(M
Y
(r)t)
n
n!
e
M
Y
(r)t
II P[(T
1
, T
2
, . . . , T
n
) A [ N
t
= n]
n
i=1
II E
II P
[e
rY
i
; Y
i
B
i
]
M
Y
(r)
.
Thus the number of claims is Poisson distributed with parameter M
Y
(r). Given
N
t
, the claim occurrence times do not change the conditional law. By Proposi-
tion C.2 the process N
t
is a Poisson process with rate M
Y
(r). The claim sizes
8. CHANGE OF MEASURE TECHNIQUES 159
are independent and have the claimed distribution. In particular
II E
Qr
[Y
n
i
] =
_
0
y
n
e
ry
dG(y)
M
Y
(r)
=
M
(n)
Y
(r)
M
Y
(r)
from which iii) follows.
iv) This follows because
Q
r
_
lim
t
C
t
t
=
t
(r)
_
= 1
and because
t
(r) is strictly increasing.
We next express the ruin probability using the measure Q
r
.
(u) = II P[ < ] = II E
Qr
[(L
r
)
1
; < ] = II E
Qr
[e
rC
e
(r)
; < ]e
ru
.
The latter is much more complicated because the joint distribution of and C
has
to be known unless (r) = 0. Let us therefore choose r = R and Q = Q
R
. First
observe that II E
Q
[C
1
u] =
t
(R) < 0. Thus the net prot condition is not fullled
and < a.s. under the measure Q. Then
(u) = II E
Q
[e
RC
; < ]e
Ru
= II E
Q
[e
RC
]e
Ru
.
Lundbergs inequality (Theorem 4.4) becomes now trivial. Because C
x [ C
0
= 0] the
descending ladder height distribution under Q. Note that (4.10) does not apply
because the limit of
x
(u) is not 0 anymore as u . Let Z(u) = II E
Q
[e
RC
[ C
0
=
u]. Then Z(u) fulls the renewal equation
Z(u) =
_
u
0
Z(u y) dB(y) +
_
u
e
R(yu)
dB(y) .
It is not dicult to show that
_
u
e
R(yu)
dB(y)
is directly Riemann integrable and its integral is
1
R
_
1
_
0
e
Ry
dB(y)
_
.
160 8. CHANGE OF MEASURE TECHNIQUES
Thus by the renewal theorem the limit of Z(u) exists and is equal to
1
R
_
0
u dB(u)
_
1
_
0
e
Ry
dB(y)
_
.
Consider rst
_
0
u dB(u). We nd
_
0
u dB(u) = II E
Q
[C
[ C
0
= 0] = II E
II P
[C
e
RC
; < [ C
0
= 0]
=
c
_
0
xe
Rx
(1 G(x)) dx =
R
2
c
(RM
t
Y
(R) (M
Y
(R) 1)) .
The integral in the numerator of the limit is
_
0
e
Ry
dB(y) = II E
Q
[e
RC
[ C
0
= 0] = II P[ < [ C
0
= 0] =
c
.
Thus
lim
u
Z(u) =
1
R
_
1
c
_
R
2
c
(RM
t
Y
(R) (M
Y
(R) 1))
=
c
(M
t
Y
(R) (M
Y
(R) 1)/R)
=
c
M
t
Y
(R) c
(8.1)
where we used the denition of the Lundberg exponent R. This proves the Cramer-
Lundberg approximation (Theorem 4.6).
8.2. The Renewal Case
In Section 5 we only considered the process at the claim times. But for the change
of the measure we need a continuous time martingale. As mentioned in Section 5
the process C
t
is not a Markov process. For our approach it is important to have
the Markov property. Thus we need to Markovize the process. The future of the
process is dependent on the time of the last claim. Let W
t
= t T
Nt
denote the time
elapsed since the last claim. Then (C
t
, W
t
) is a Markov process. Alternatively we
can consider V
t
= T
Nt+1
t the time left till the next claim. This seems a little bit
strange because V
t
is not observable. But recall that in the Ammeter model we
worked with a non-observable stochastic process which was useful in order to obtain
Lundbergs inequality in the general case. The stochastic process (C
t
, V
t
) is also
a Markov process. But note that the natural ltrations of C
t
and of (C
t
, V
t
) are
dierent.
8. CHANGE OF MEASURE TECHNIQUES 161
8.2.1. Markovization Via the Time Since the Last Claim
In order to construct a martingale we have to assume that the interarrival time
distribution F is absolutely continuous with density F
t
(t). The following martingale
follows from the theory of piecewise deterministic Markov processes, see [66].
Lemma 8.2. Assume that M
Y
(r) < such that the unique solution (r) to
M
Y
(r)M
T
( cr) = 1
exists (such a solution exists if r 0). Then the process
L
r
t
= e
r(Ctu)
M
Y
(r)
e
((r)+cr)Wt
1 F(W
t
)
_
Wt
e
((r)+cr)s
F
t
(s) ds e
(r)t
is a martingale.
Remark. Considering the above martingale at the times 0, T
1
, . . . only yields the
martingale obtained in Lemma 5.1.
Proof. See Dassios and Embrechts [28, Thm.10].
Dene now the measures for A T
t
Q
r
[A] = II E
II P
[L
r
t
; A] .
Lemma 8.1 remains true also in the renewal case. Also here, we suppose that T
t
F(t) = M
Y
(r)
_
t
0
e
((r)+cr)s
dF(s). In particular
lim
t
1
t
(C
t
u) =
t
(r)
Q
r
-a.s..
162 8. CHANGE OF MEASURE TECHNIQUES
iv) If r ,= 0 then Q
r
and P are singular on the whole -algebra T.
Proof. i), ii) and iv) are proved as in Lemma 8.1.
iii) Let n IIN and A
1
, A
2
, . . . , A
n
, B
1
, B
2
, . . . , B
n
be Borel sets. As long as we did
not prove that the measure Q
r
can be extended to T we have to choose A
1
, . . . , A
n
such that T
n
remains nite. Then because W
Tn
= 0
Q
r
[T
1
A
1
, T
2
T
1
A
2
, . . . , T
n
T
n1
A
n
, Y
1
B
1
, Y
2
B
2
, . . . , Y
n
B
n
]
= II E
II P
[L
r
Tn
; T
1
A
1
, T
2
T
1
A
2
, . . . , T
n
T
n1
A
n
,
Y
1
B
1
, Y
2
B
2
, . . . , Y
n
B
n
]
= II E
II P
_
exp
_
r
n
i=1
Y
i
_
exp
_
((r) +cr)
n
i=1
(T
i
T
i1
)
_
; T
1
A
1
,
T
2
T
1
A
2
, . . . , T
n
T
n1
A
n
, Y
1
B
1
, Y
2
B
2
, . . . , Y
n
B
n
_
=
n
i=1
II E
II P
[e
rY
i
; Y
i
B
i
]II E
II P
[e
((r)+cr)(T
i
T
i1
)
; (T
i
T
i1
) A
i
] .
Thus (C
t
, W
t
) is the renewal risk model with the parameters given by the assertion.
A straightforward calculation yields
lim
t
1
t
(C
t
u) = c
=
t
(r)
Q
r
-a.s..
The ruin probability can be expressed as
(u) = II E
Qr
_
e
rC
(1 F(W
))e
((r)+cr)W
M
Y
(r)
_
W
e
((r)+cr)s
F
t
(s) ds
e
(r)
; <
_
e
ru
= II E
Qr
[e
rC
e
(r)
; < ]e
ru
where we used that W
F(t) = M
Y
(r)
_
t
0
e
((r)+cr)s
dF(s). In particular
lim
t
1
t
(C
t
u) =
t
(r)
Q
r
-a.s..
iv) If r ,= 0 then Q
r
and P are singular on the whole -algebra T.
Proof. Left as an exercise.
164 8. CHANGE OF MEASURE TECHNIQUES
For the ruin probability we obtain
(u) = (M
Y
(r))
1
II E
Qr
[e
rC
e
((r)+cr)V
e
(r)
; < ]e
ru
= II E
Qr
[e
rC
e
(r)
; < ]e
ru
.
Note that V
is independent of T
t
and the intensity
levels by
L
i
. From the lemma below it follows again that Q
r
can be extended to the
whole -algebra T
t
.
Lemma 8.6.
i) Let t > s and A T
s
T
t
. Then II E
II P
[L
r
t
; A] = II E
II P
[L
r
s
; A].
ii) Let T be a stopping time and A T
T
such that A T < . Then
Q
r
[A] = II E
II P
[L
r
T
; A] .
iii) (C
t
,
t
, V
t
) is an Ammeter risk model under the measure Q
r
with claim size
distribution
G(x) =
_
x
0
e
ry
dG(y)/M
Y
(r), claim arrival intensity
t
= M
Y
(r)
t
,
L
i
= M
Y
(r)L
i
and
H() = e
(cr+(r))
_
0
e
l(M
Y
(r)1)
dH(l). In particular,
II E
Qr
[C
u] =
t
(r).
iv) If r ,= 0 then Q
r
and P are singular on the whole -algebra T.
Proof. i), ii) and iv) are proved as in Lemma 8.1.
iii) We rst want to show that the processes (C
t
C
(i1)
: (i 1) t < i)
are independent and follow the same law. Let A
i
be a set measurable with respect
to (L
i
, N
i
N
(i1)
, T
N
(i1)
+1
, . . . , T
N
i
, Y
N
(i1)
+1
, . . . , Y
N
i
). Then for n IIN
Q
r
[A
1
, . . . , A
n
] = II E
II P
[e
r(C
n
u)
e
(L
n+1
(M
Y
(r)1)cr(r))
e
(r)n
; A
1
, . . . , A
n
]
= II E
II P
[e
r(C
n
u)
e
(r)n
; A
1
, . . . , A
n
]
=
n
i=1
e
(r)
II E
II P
[e
r(C
i
C
(i1)
)
; A
i
] .
Next we show that the claim sizes are iid., independent of N
t
and have the required
distribution. Let n IIN and A be a (L
1
, T
1
, . . . , T
n
)-measurable set. Let B
i
be
Borel sets. Then
Q
r
[N
= n, A, Y
1
B
1
, . . . , Y
n
B
n
]
= e
(r)
II E
II P
[e
r(C
u)
; N
= n, A, Y
1
B
1
, . . . , Y
n
B
n
]
= e
(cr+(r))
II P[N
= n, A]
n
i=1
II E
II P
[e
rY
i
; Y
i
B
i
] .
166 8. CHANGE OF MEASURE TECHNIQUES
Next we verify the distribution of L
1
and N
= n] = e
(cr+(r))
M
Y
(r)
n
_
A
()
n
n!
e
dH()
=
_
A
(M
Y
(r))
n
n!
e
M
Y
(r)
e
(cr+(r))
e
(M
Y
(r)1)
dH() .
Thus L
1
and N
= n then times (T
1
, . . . , T
n
) have the same distribution as the order
statistics of n iid. uniformly in (0, ) distributed random variables (Proposition C.2),
i.e. the same conditional distribution as under II P. Let n IIN, A be a (T
1
, . . . , T
n
)-
measurable set and B be a Borel set such that II P[L
1
B] ,= 0. Then
Q
r
[A [ L
1
B, N
= n] =
e
(cr+(r))
II P[A, L
1
B, N
= n]M
Y
(r)
n
e
(cr+(r))
II P[L
1
B, N
= n]M
Y
(r)
n
= II P[A [ L
1
B, N
= n] = II P[A [ N
= n] .
The last assertion follows straightforwardly noting that
II E
Qr
[C
u] = (c II E
Qr
[M
Y
(r)L
i
]II E
Qr
[Y
i
]).
, V
) as u ? In
order to use the approach which was successful in the Cramer-Lundberg and in the
renewal case this distribution has to be known. Anyway, with a slightly modied
renewal approach it is possible to prove the Cramer-Lundberg approximation, see
[71] for details.
In order to nd a lower bound for the ruin probability dene
1
:= infk : k
IIN, C
k
< 0 the rst time when the process becomes negative at a regeneration
time. We assume now that there exists an r > R such that M
L
((M
Y
(r)1)) < .
Obviously II P[ < ] II P[
1
< ] and
II P[
1
< ] = II E
Q
[e
RC
1
e
(
1
(M
Y
(R)1)cR)V
1
]e
Ru
= II E
Q
[e
RC
1
]e
Ru
e
RII E
Q
[C
1
]
e
Ru
.
8. CHANGE OF MEASURE TECHNIQUES 167
Because II E
Q
[e
(rR)(Ctu)
] exists all moments of the increments of the random walk
C
k
u exist. Using the renewal theorem one can show that II E
Q
[C
1
] converges to
a nite value as u .
The case with a general F
1
(, v) is left to the reader as an exercise.
Bibliographical Remarks
The change of measure method is frequently used in the literature without men-
tioning the connection to martingales, see for instance [6], [7] or [8]. The technique
used in this section can also be found in [9] or [66]. Lemma 8.4 was used in an early
version of [27], but does not appear in the nal version.
168 9. THE MARKOV MODULATED RISK MODEL
9. The Markov Modulated Risk Model
In the Ammeter model the claim arrival intensity was constant during the interval
((k 1), k). For instance was one year. It would be preferable, if the intensity
was allowed to change more frequently. This can be achieved by choosing smaller.
But it would also be preferable that dierent intensities levels could be holding for
time intervals of dierent lengths. A model fullling the above requirements is the
Markov modulated risk model. One rst constructs a continuous time Markov chain
on a nite state space. This Markov chain represents an environment process for
the risk business. The environment determines the intensity level. Moreover, also
the claim size distribution can dier for dierent states of the environment.
9.1. Denition of the Markov Modulated Risk Model
Let J
t
be a continuous time Markov chain with state space (1, 2, . . . , ) and
intensity matrix = (
ij
), where IIN 0 and
ij
0 for i ,= j and
ii
=
j,=i
ij
. This means that, if J
t
= i, then J
t+s
= i for a Exp(
ii
) distributed
time and the next state will be j ,= i with probability
ij
/(
ii
). In order to obtain
an ergodic model we assume that
ii
< 0 for all i . Let L
i
: i denote
the intensity levels and let G
i
(y) : i be distribution functions of positive
random variables. L
i
will be the claim arrival intensity and G
i
(y) will be the claim
size distribution during the time where the environment process is in state i. For a
formal denition let
0
= 0 and for n IIN
n+1
= inft >
n
: J
t
,= J
t
0
(1 G
i
(y)) dy be
the mean value and M
(i)
Y
(r) =
_
0
e
ry
dG
i
(y) be the moment generating function of
a claim in state i. By M
Y
(r) = maxM
(i)
Y
(r) : i we denote the maximal value
of the moment generating functions. Denote by I the identity matrix. For r such
that M
Y
(r) < let S(r) be the diagonal matrix with S
ii
(r) = L
i
(M
(i)
Y
(r) 1) and
(r) := +S(r) crI.
Denition 9.1. Let A be a square matrix. Then we denote by e
tA
the matrix
e
tA
:=
n=0
t
n
n!
A
n
.
Lemma 9.2. For the Markov modulated risk model C
t
one has
II E[e
r(Ctu)
1I
Jt=j
[ J
0
= i] = (e
t(r)
)
ij
provided M
Y
(r) < . In particular II P[J
t
= j [ J
0
= i] = (e
t
)
ij
.
Proof. Let f
ij
(t) = II E[e
r(Ctu)
1I
Jt=j
[ J
0
= i]. For a small time interval we have
II P[N
t+h
N
t
= 0 [ J
t
= k] = 1L
k
h+o(h), II P[N
t+h
N
t
= 1 [ J
t
= k] = L
k
h+o(h)
and
II P[J
t+h
= l [ J
t
= k] = I
kl
+
kl
h +o(h) .
Therefore, for h small,
f
ij
(t +h) = f
ij
(t)(1 +
jj
h +o(h))
_
(1 L
j
h +o(h))e
rch
+ (L
j
h +o(h))e
rch
M
(j)
Y
(r)
_
+
k=1
k=j
f
ik
(t)(
kj
h +o(h))e
rch
+o(h)
= f
ij
(t)e
rch
+f
ij
(t)L
j
h(M
(j)
Y
(r) 1) +
k=1
f
ik
(t)
kj
h +o(h) .
Thus f
ij
(t) fulls the dierential equation
f
t
ij
(t) = crf
ij
(t) +f
ij
(t)L
j
(M
(j)
Y
(r) 1) +
k=1
f
ik
(t)
kj
=
k=1
f
ik
(t)
kj
(r)
with initial condition f
ij
(0) = I
ij
. The only solution to the above system of dier-
ential equations is f
ij
(t) = (e
t(r)
)
ij
.
170 9. THE MARKOV MODULATED RISK MODEL
Of special interest is the case where J
0
is distributed such that J
t
becomes
stationary. The initial distribution has to full
e
t
=
where we write as a row vector. The condition is equivalent to = 0. We
assume that the Markov chain is irreducible, which means that is unique and
i
> 0 for all i .
Let for the moment u = 0 and let V
i
(t) =
_
t
0
1I
Js=i
ds denote the time, J
s
has
spent in state i. Then
C
t
t
=
i=1
V
i
(t)
t
_
t
0
1I
Js=i
dC
s
V
i
(t)
where
_
t
0
1I
Js=i
dC
s
=
_
t
0
1I
Js=i
c ds
Nt
j=1
1I
J
T
j
=i
Y
i
.
It is well-known (and follows from the law of large numbers) that V
i
(t)/t
i
. The
random variable
_
t
0
1I
Js=i
dC
s
has the same distribution as C
(i)
V
i
(t)
and it follows
from the classical model that C
(i)
V
i
(t)
/V
i
(t) tends to c L
i
i
a.s.. Thus C
t
/t converges
a.s. to c
i=1
i
L
i
i
. In order to avoid (u) = 1 for all u the latter limit should
be strictly positive. This means the net prot condition reads
c >
i=1
i
L
i
i
.
9.2. The Lundberg Exponent and Lundbergs Inequality
The matrix e
(r)
has only strictly positive elements by Lemma 9.2. Let e
(r)
=
spr e
(r)
denote the largest absolute value of the eigenvalues. By the Frobenius the-
orem e
(r)
is an eigenvalue of e
(r)
, it is the only algebraic eigenvalue with absolute
value spr e
(r)
and its eigenvector g(r) consists of strictly positive elements. More-
over g(r) is the only non-trivial eigenvector consisting of only positive elements. It
follows in particular that g(0) = 1 because 1 = 0. We normalize g(r) such that
g(r) = 1. We nd the following martingale.
9. THE MARKOV MODULATED RISK MODEL 171
Lemma 9.3. The stochastic process
L
r
t
= g
Jt
(r)/II E[g
J
0
(r)] e
r(Ctu)(r)t
is a martingale. Moreover the function (r) is convex and
t
(0) =
_
c
i=1
i
L
i
i
_
< 0 .
Proof. In order to simplify the notation we write g
i
instead of g
i
(r). Note that
II E[g
Jt
e
r(Ctu)
[ J
0
= i] = (e
t(r)
g)
i
= e
(r)t
g
i
.
Because the increments of C
t+s
: s 0 depend on T
t
via J
t
only, the martingale
property is proved.
It follows from Lemma 1.9 that log
i=1
II E[e
r(Ctu)
; J
t
= j [ J
0
= i] is strictly
convex. It is clear that
lim
n
((e
(r)
)
n
1)
1/n
i
e
(r)
.
Denote by g = maxg
i
. Then
lim
n
((e
(r)
)
n
1)
1/n
i
= lim
n
((e
(r)
)
n
g1)
1/n
i
lim
n
((e
(r)
)
n
g)
1/n
i
= e
(r)
.
Thus there exists a series f
n
(r) of strictly convex functions converging to (r). Let
r
1
< r
2
such that M
Y
(r
2
) < . Let (0, 1). Then
(r
1
+ (1 )r
2
) = lim
n
f
n
(r
1
+ (1 )r
2
)
lim
n
f
n
(r
1
) + (1 )f
n
(r
2
) = (r
1
) + (1 )(r
2
)
and (r) is convex.
Note that (r) is an eigenvalue of (r), i.e. a solution to the equation det((r)
I) = 0. By the implicit function theorem (r) is dierentiable and therefore also
g(r) is dierentiable. We have (r)g(r) = (r)g(r) = (r). In particular
t
(r) =
t
(r)g(r) +(r)g
t
(r) .
Letting r = 0 the result for
t
(0) follows.
Because (r) is convex there might exist a second solution R ,= 0 to the equation
(r) = 0. If such a solution exists then R > 0 and the solution is unique. We call it
again Lundberg exponent or adjustment coecient. Let us now assume that
R exists. As in Chapter 8 let us dene the new measure Q
r
[A] = II E
II P
[L
r
t
; A]. It will
again follow from the lemma below that Q
r
can be extended to T.
172 9. THE MARKOV MODULATED RISK MODEL
Lemma 9.4.
i) Let t > s and A T
s
T
t
. Then II E
II P
[L
r
t
; A] = II E
II P
[L
r
s
; A].
ii) Let T be a stopping time and A T
T
such that A T < . Then
Q
r
[A] = II E
II P
[L
r
T
; A] .
iii) Under the measure Q
r
, (C
t
, J
t
) is a Markov modulated risk model with in-
tensity matrix
ij
= (g
i
(r))
1
g
j
(r)
ij
if i ,= j, claim size distributions
G
i
(x) =
_
x
0
e
ry
dG
i
(y)/M
(i)
Y
(r) and claim arrival intensities
L
i
= L
i
M
(i)
Y
(r). In particular
lim
t
1
t
(C
t
u) =
t
(r)
Q
r
-a.s..
Proof. i) and ii) are proved as in Lemma 8.1.
iii) In order to simplify the notation we write g
i
instead of g
i
(r). Let B
k
be (
k
k1
, N
k
N
k1
, T
j
T
j1
, Y
j
; N
k1
< j N
k
)-measurable sets, n IIN and
i
k
. Before one has proved that Q
r
can be extended to T one has to choose sets
B
k
such that we only consider the process in nite time. Let i
k
(1, . . . , ). Then
Q
r
[J
k1
= i
k
, J
n
= i
n+1
, B
k
, 1 k n]
=
g
i
n+1
II E
II P
[g
J
0
]
II E
II P
_
n
k=1
e
r(C
k
C
k1
)
e
(r)(
k
k1
)
;
J
k1
= i
k
, J
n
= i
n+1
, B
k
, 1 k n
_
=
g
i
n+1
II E
II P
[g
J
0
]
II E
II P
[e
r(C
1
u)
e
(r)
1
; J
0
= i
1
, B
1
]
k=2
II E
II P
[e
r(C
k
C
k1
)
e
(r)(
k
k1
)
; J
k1
= i
k
, B
k
[ J
k2
= i
k1
]
II P[J
n
= i
n+1
[ J
n1
= i
n
]
and therefore the process in the interval [
k
,
k+1
) depends on T
via J
k1
only.
In particular
Q
r
[J
k
= i
k+1
, B
k
, 1 k n [ J
0
= i
1
]
=
g
i
n+1
g
i
1
II E
II P
[e
r(C
1
u)
e
(r)
1
; B
1
[ J
0
= i
1
]
k=2
II E
II P
[e
r(C
k
C
k1
)
e
(r)(
k
k1
)
; J
k1
= i
k
, B
k
[ J
k2
= i
k1
]
II P[J
n
= i
n+1
[ J
n1
= i
n
]
9. THE MARKOV MODULATED RISK MODEL 173
and
Q[B
2
[ J
0
= i
1
, J
1
= i
2
] = II E
II P
[e
r(C
2
C
1
)
e
(r)(
2
1
)
; B
2
[ J
0
= i
1
, J
1
= i
2
]
= II E
II P
[e
r(C
2
C
1
)
e
(r)(
2
1
)
; B
2
[ J
1
= i
2
] = Q[B
2
[ J
1
= i
2
]
and therefore the process in the interval [
n
,
n+1
) depends on T
n
via J
n
only.
Next we compute the distribution of
1
.
Q[
1
> v [ J
0
= i] = II E
II P
[e
r(Cvu)(r)v
;
1
> v [ J
0
= i]
= e
(L
i
(M
(i)
Y
(r)1)cr(r))v
e
ii
v
and therefore
1
is Exp(((r) (r)I)
ii
) distributed. It follows that J
t
is a
continuous time Markov chain under Q
r
. In order to nd the intensity we compute
Q[J
1
= j [ J
0
= i] =
g
j
II P[J
1
= j [ J
0
= i]
g
i
II E
II P
[e
(L
i
(M
(i)
Y
(r)1)cr(r))
1
[ J
0
= i]
=
g
j
ij
g
i
((r) (r)I)
ii
and the desired intensity matrix under Q
r
follows. Note that we should have con-
sidered Q[J
1
= j [ J
0
= i,
1
t] which would have given the same result.
Let n IIN, A be a (T
1
, . . . , T
n
)-measurable set and B
k
be Borel sets. Then
Q[
1
> v, N
v
= n, A, Y
k
B
k
, 1 k n [ J
0
= i]
= e
(cr+(r))v
II P[
1
> v, N
v
= n, A [ J
0
= i]
n
k=1
_
B
k
e
ry
dG
i
(y)
and thus the claim size distribution is as desired. Moreover
Q[A [ N
v
= n,
1
> v, J
0
= i] = II P[A [ N
v
= n,
1
> v, J
0
= i]
from which it follows that, conditioned on the number of claims, the claim times are
distributed as in a Poisson process. Finally
Q[N
v
= n [
1
> v, J
0
= i] =
(L
i
M
(i)
Y
(r)v)
n
n!
e
L
i
M
(i)
Y
(r)v
which shows that (C
t
, J
t
) is a Markov modulated risk model.
Under the measure Q
r
the process
L
s
t
=
g
Jt
(r +s)
g
Jt
(r)
e
s(Ctu)
e
((s+r)(r))t
174 9. THE MARKOV MODULATED RISK MODEL
is a martingale because for t v
II E
Qr
_
g
Jt
(r +s)
g
Jt
(r)
e
s(Ctu)((s+r)(r))t
T
v
_
=
II E
II P
_
g
J
t
(r+s)
g
J
t
(r)
e
s(Ctu)((s+r)(r))t
g
J
t
(r)
II E
II P
[g
J
0
(r)]
e
r(Ctu)(r)t
T
v
_
g
Jv
(r)
II E
II P
[g
J
0
(r)]
e
r(Cvu)(r)v
= II E
II P
_
g
Jt
(r +s)
g
Jv
(r)
e
(r+s)(Ctu)(r+s)t
T
v
_
e
r(Cvu)+(r)v
=
g
Jv
(r +s)
g
Jv
(r)
e
(r+s)(Cvu)(r+s)v
e
r(Cvu)+(r)v
=
L
s
v
.
As in Lemma 9.3 it follows that
lim
t
1
t
(C
t
u) =
d
ds
((r +s) (r))[
s=0
=
t
(r) .
Let again denote the time of ruin and (u) denote the ruin probability. If we
write (u) in terms of the measure Q
r
we obtain
(u) = II E
II P
[g
J
0
(r)]II E
Qr
[(g
J
(r))
1
e
rC
e
(r)
; < ]e
ru
. (9.1)
The latter expression is only simpler than II P[ < ] if r = R. In the sequel we
write Q instead of Q
R
and g
i
instead of g
i
(R).
We now can prove Lundbergs inequality.
Theorem 9.5. Let (C
t
, J
t
) be a Markov modulated risk model and assume that
R exists. Let g
i
: 1 i be as above and g
min
= ming
i
: 1 i . Then
(u) < II E
II P
[g
J
0
]/g
min
e
Ru
.
Proof. Note that g
min
> 0. Thus the assertion follows readily from (9.1) noting
that e
RC
< 1.
9. THE MARKOV MODULATED RISK MODEL 175
9.3. The Cramer-Lundberg Approximation
For proving the Cramer-Lundberg approximation in the classical case (8.1) and
in the renewal case it was no problem to use the renewal equation because each
ladder time was a regeneration point. This is not the case anymore for the Markov
modulated risk model. We therefore have to nd a slightly dierent version of the
ladder epochs.
Assume that L
1
,= 0 and that J
0
= 1. Let
:= inft > 0 : J
t
= 1 and C
t
= inf
0st
C
s
.
Because of the lack of memory property of the exponential distribution
is a
regeneration point. Note that Q[
< ] = 1. If C
< 0 [ C
0
= u, J
0
= 1] .
In order to show that Z(u) converges as u we have to show that z(u) is
directly Riemann integrable. To avoid lim
u
Z(u) = 0 we assume that there exists
an r > R such that M
Y
(r) < .
Let us rst show that z(u) is a continuous function. Dene
z
i
(u) = II E
Q
[(g
J
)
1
e
RC
; C
< 0 [ C
0
= u, J
0
= i] .
Note that z
i
(u) is bounded. Then for h > 0 small we condition on time T h where
T = T
1
1
and T
1
is the time of the rst claim. Note that T is under the measure
Q exponentially distributed with parameter
L
1
11
. Then
z(u) = e
(
L
1
11
)h
z(u +ch) +
_
h
0
_
0
z(u +ct y) d
G
1
(y)
L
1
e
(
L
1
11
)t
dt
+
j=2
_
h
0
z
j
(u +ct)
1j
e
(
L
1
11
)t
dt .
176 9. THE MARKOV MODULATED RISK MODEL
Letting h 0 shows that z(u) is right continuous. Replacing u by u ch gives
z(u ch) = e
(
L
1
11
)h
z(u) +
_
h
0
_
0
z(u c(h t) y) d
G
1
(y)
L
1
e
(
L
1
11
)t
dt
+
j=2
_
h
0
z
j
(u c(h t))
1j
e
(
L
1
11
)t
dt .
and z(u) is left continuous too.
It remains to show that z(u) has a directly Riemann integrable upper bound.
Then it will follow from Lemma C.11 that z(u) is directly Riemann integrable. It is
clear that
g
min
z(u) Q[C
< 0 [ C
0
= u, J
0
= 1] = Q[C
> u [ J
0
= 1, C
0
= 0] .
The latter function is monotone. Thus it is enough to show that it is integrable.
But
_
0
Q[C
> u [ J
0
= 1, C
0
= 0] du = II E
Q
[C
[ J
0
= 1, C
0
= 0] < .
The latter follows from the Wiener-Hopf factorization because
II E
Q
[e
(rR)/2 (C
)
[ J
0
= 1, C
0
= 0]
= II E
II P
[e
(r+R)/2 (C
)
;
< [ J
0
= 1, C
0
= 0] <
since II E
II P
[e
rCt
] < for all t 0. It follows readily that for
= infT
n
: n
0, J
Tn
= 1 we also have II E
II P
[exprC
Z(u) = II E
Q
[(g
J
)
1
e
RC
[ C
0
= u]
9. THE MARKOV MODULATED RISK MODEL 177
the analogue of Z(u). Then
Z(u) =
_
u
Z(u y) dB
1
(y) + z(u)
where
z(u) = II E
Q
[(g
J
)
1
e
RC
; [ C
0
= u] .
Because Q[ [ C
0
= u] tends to 0 as u we nd that z(u) tends to 0 as
u . Because Z(x) is a bounded function we can interchange the limit and the
integration to nd that
lim
u
Z(u) =
_
lim
u
Z(u y) dB
1
(y) = C .
1
_
x
0
(1 G
1
(y)) dy
is subexponential. Let K =
i
L
i
K
i
. Then
lim
u
(u)
_
u
(1 G
1
(y)) dy
=
K
c
i
L
i
i
independent of the distribution of the initial state.
178 9. THE MARKOV MODULATED RISK MODEL
Let us consider the problem intuitively. Assume the process starts in state 1. Con-
sider the random walk generated by the process by considering only the time points
when the process J
t
jumps to state 1. The increments of this random walk have a
left tail that is asymptotically comparable with the tail of G
1
(only the heaviest tail
contributes). Thus it follows as in the renewal case that the ruin probability of the
random walk has the desired behaviour.
The time between two regeneration points has a light tailed distribution. The
dierence between the inmum of the random walk and the inmum of the risk pro-
cess is smaller than c times the distance between the corresponding two regeneration
points. The inmum of the random walk has a tail proportional to
_
u
(1G
1
(y)) dy
which will asymptotically dominate the light tail of the distance between two regen-
eration points. Thus the tail of the distribution of the inmum of the random walk
and tail of the distribution of the inmum of the risk process are asymptotically
equivalent.
9.5. Finite Time Lundberg Inequalities
Let 0 y < y < . As for the other models we want to consider probabilities of the
form II P[yu < yu]. We can proceed exactly as in the classical case (Section 4.14)
II E
II P
[g
J
0
(r)]e
ru
= II E
II P
[g
J
yu
(r)e
rC
yu
e
(r)(yu)
]
> II E
II P
[g
J
(r)e
rC
e
(r)
; yu < yu]
> g
min
(r)II E
II P
[e
(r)
[ yu < yu]II P[yu < yu]
g
min
(r)e
maxy(r),y(r)u
II P[yu < yu] .
Hence we obtain the inequality
II P[yu < yu] <
II E
II P
[g
J
0
(r)]
g
min
(r)
e
minry(r),ry(r)u
.
Dening the nite time Lundberg exponent
R(y, y) = supminr y(r), r y(r) : r IR
we obtain the nite time Lundberg inequality
II P[yu < yu] <
II E
II P
[g
J
0
( r)]
g
min
( r)
e
R(y,y)u
where r is the argument maximizing minr y(r), r y(r). In the discussion
of R(y, y) in Section 4.14 we only used to property that (r) is convex. Thus
9. THE MARKOV MODULATED RISK MODEL 179
the same discussion applies. Let r
y
be the argument maximizing r y(r) and
let y
0
= (
t
(R))
1
denote the critical value. Then we obtain the two nite time
Lundberg inequalities
II P[0 < yu [ C
0
= u] <
II E
II P
[g
J
0
(r
y
)]
g
min
(r
y
)
e
R(0,y)u
where R(0, y) > R if y < y
0
, and
II P[yu < < [ C
0
= u] <
II E
II P
[g
J
0
(r
y
)]
g
min
(r
y
)
e
R(y,)u
where R(y, ) > R if y > y
0
.
Let r
= supr IR : M
Y
(r) < . The next result can be proved in the same
way as Theorem 4.24 was proved.
Theorem 9.8. Assume that R < r
. Then
u
y
0
in probability on the set < .
Bibliographical Remarks
The Markov modulated risk model was introduced as a semi-Markov model by
Janssen [56]. The useful formulation as a model in continuous time goes back to
Asmussen [8], where also Lemma 9.2, Theorem 9.5 and Theorem 9.6 were proved.
A version of Theorem 9.7, where the limiting constant was dependent on the initial
distribution, can be found in [10]. The Theorem presented here is proved in [12].
The model is also discussed in [66].
180 A. STOCHASTIC PROCESSES
A. Stochastic Processes
We start with some denitions.
Denition A.1. Let I be either IIN or [0, ). A stochastic process on I with
state space E is a family of random variables X
t
: t I on E. Let E be a
topological space. A stochastic process is called cadlag if it has a.s. right continuous
paths and the limits from the left exist. It is called continuous if its paths are
a.s. continuous. We will often identify a cadlag (continuous, respectively) stochastic
process with a random variable on the space of cadlag (continuous) functions. For
the rest of these notes we will always assume that all stochastic processes on [0, )
are cadlag.
Denition A.2. A cadlag stochastic process N
t
on [0, ) is called point pro-
cess if a.s.
i) N
0
= 0,
ii) N
t
N
s
for all t s and
iii) N
t
IIN for all t (0, ).
We denote the jump times by T
1
, T
2
, . . ., i.e. T
k
= inft 0 : N
t
k for all k IIN.
In particular T
0
= 0. A point process is called simple if T
0
< T
1
< T
2
< .
Denition A.3. A stochastic process X
t
is said to have independent incre-
ments if for all n IIN and all real numbers 0 = t
0
< t
1
< t
2
< < t
n
the random
variables X
t
1
X
t
0
, X
t
2
X
t
1
, . . . , X
tn
X
t
n1
are independent. A stochastic pro-
cess is said to have stationary increments if for all n IIN and all real numbers
0 = t
0
< t
1
< t
2
< < t
n
and all h > 0 the random vectors (X
t
1
X
t
0
, X
t
2
X
t
1
, . . . , X
tn
X
t
n1
) and (X
t
1
+h
X
t
0
+h
, X
t
2
+h
X
t
1
+h
, . . . , X
tn+h
X
t
n1
+h
) have
the same distribution.
Denition A.4. An increasing family T
t
of -algebras is called ltration if
T
t
T for all t. A stochastic process X
t
is called T
t
-adapted if X
t
is T
t
-
measurable for all t 0. The natural ltration T
X
t
of a stochastic process X
t
is the smallest ltration such that the process is adapted. If a stochastic process is
considered then we use its natural ltration if nothing else is mentioned. A ltration
is called right continuous if T
t+
:=
s>t
T
s
= T
t
for all t 0.
A. STOCHASTIC PROCESSES 181
Denition A.5. A T
t
-stopping time is a random variable T on [0, ] or IIN
such that T t T
t
for all t 0. If we stop a stochastic process then we
usually assume that T is a stopping time with respect to the natural ltration of the
stochastic process. The -algebra
T
T
:= A T : A T t T
t
for all t 0
is called the information up to the stopping time T.
A.1. Bibliographical Remarks
For further literature on stochastic processes see for instance [30], [39], [58], [59],
[67] or [88].
182 B. MARTINGALES
B. Martingales
Denition B.1. A stochastic process M
t
with state space IR is called a T
t
-
martingale (-submartingale, -supermartingale, respectively) if
i) M
t
is adapted,
ii) II E[M
t
] exists for all t 0,
iii) II E[M
t
[ T
s
] = (, ) M
s
a.s. for all t s 0.
We simply say M
t
is a martingale (submartingale, supermartingale) if it is a
martingale (submartingale, supermartingale) with respect to its natural ltration.
The following two propositions are very important for dealing with martingales.
Proposition B.2. (Martingale stopping theorem) Let M
t
be a T
t
-
martingale (-submartingale, -supermartingale, respectively) and T be a T
t
-stop-
ping time. Assume that T
t
is right continuous. Then also the stochastic process
M
Tt
: t 0 is a T
t
-martingale (-submartingale, -supermartingale). Moreover,
II E[M
t
[ T
T
] = (, )M
Tt
.
Proof. See [30, Thm. 12], [39, Thm. II.2.13] or [66, Thm. 10.2.5].
Proposition B.3. (Martingale convergence theorem) Let M
t
be a T
t
-
martingale such that lim
t
II E[M
t
] < (or equivalently sup
t0
II E[[M
t
[] < ). If T
t
:= lim
t
M
t
exists a.s. and is inte-
grable.
Proof. See for instance [30, Thm. 6] or [66, Thm. 10.2.2].
Note that in general
II E[M
] = II E[ lim
t
M
t
] ,= lim
t
II E[M
t
] = II E[M
0
] .
B.1. Bibliographical Remarks
The theory of martingales can also be found in [30], [39] or [66].
C. RENEWAL PROCESSES 183
C. Renewal Processes
C.1. Poisson Processes
Denition C.1. A point process N
t
is called (homogeneous) Poisson process
with rate if
i) N
t
has stationary and independent increments,
ii) II P[N
h
= 0] = 1 h +o(h) as h 0,
iii) II P[N
h
= 1] = h +o(h) as h 0.
Remark. It follows readily that II P[N
h
2] = o(h) and that the point process is
simple.
We give now some alternative denitions of the Poisson process.
Proposition C.2. Let N
t
be a point process. Then the following are equivalent:
i) N
t
is a Poisson process with rate .
ii) N
t
has independent increments and N
t
Pois(t) for each xed t 0.
iii) The interarrival times T
k
T
k1
: k 1 are independent and Exp() dis-
tributed.
iv) For each xed t 0, N
t
Pois(t) and given N
t
= n the occurrence points
have the same distribution as the order statistics of n independent uniformly on
[0, t] distributed points.
v) N
t
has independent increments such that II E[N
1
] = and given N
t
= n the
occurrence points have the same distribution as the order statistics of n indepen-
dent uniformly on [0, t] distributed points.
vi) N
t
has independent and stationary increments such that II P[N
h
2] = o(h)
and II E[N
1
] = .
184 C. RENEWAL PROCESSES
Proof. i) = ii) Let p
n
(t) = II P[N
t
= n]. Then show that p
n
(t) is continuous
and dierentiable. Finding the dierential equations and solving them shows ii).
The details are left as an exercise.
ii) = iii) We rst show that N
t
has stationary increments. It is enough to
show that N
t+h
N
h
is Pois(t) distributed. For the moment generating functions
we obtain
e
(t+h)(e
r
1)
= II E[e
r(N
t+h
N
h
)
e
rN
h
] = II E[e
r(N
t+h
N
h
)
]e
h(e
r
1)
.
It follows that
II E[e
r(N
t+h
N
h
)
] = e
t(e
r
1)
.
This proves that N
t+h
N
h
is Pois(t) distributed.
Let t
0
= 0 s
1
< t
1
s
2
< t
2
s
n
< t
n
. Then
II P[s
k
< T
k
t
k
, 1 k n]
= II P[N
s
k
N
t
k1
= 0, N
t
k
N
s
k
= 1, 1 k n 1,
N
sn
N
t
n1
= 0, N
tn
N
sn
1]
= e
(snt
n1
)
(1 e
(tnsn)
)
n1
k=1
e
(s
k
t
k1
)
(t
k
s
k
)e
(t
k
s
k
)
= (e
sn
e
tn
)
n1
n1
k=1
(t
k
s
k
)
=
_
t
1
s
1
_
tn
sn
n
e
yn
dy
n
dy
1
=
_
t
1
s
1
_
t
2
z
1
s
2
z
1
_
tnz
1
z
n1
snz
1
z
n1
n
e
(z
1
++zn)
dz
n
dz
1
.
It follows that the joint density of T
1
, T
2
T
1
, . . . , T
n
T
n1
is
n
e
(z
1
++zn)
and therefore they are independent and Exp() distributed.
iii) = iv) Note that T
n
is (n, ) distributed. Thus II P[N
t
= 0] = II P[T
1
> t] =
e
t
and for n 1
II P[N
t
= n] = II P[N
t
n] II P[N
t
n + 1] = II P[T
n
t] II P[T
n+1
t]
=
_
t
0
n
s
n1
(n 1)!
e
s
ds
_
t
0
n+1
s
n
n!
e
s
ds
=
_
t
0
d
ds
(s)
n
n!
e
s
ds =
(t)
n
n!
e
t
.
C. RENEWAL PROCESSES 185
The joint density of T
1
, . . . , T
n+1
is for t
0
= 0
n+1
k=1
e
(t
k
t
k1
)
=
n+1
e
t
n+1
.
Thus the joint conditional density of T
1
, . . . , T
n
given N
t
= n is
_
t
n+1
e
t
n+1
dt
n+1
_
t
0
_
t
t
1
_
t
t
n1
_
t
n+1
e
t
n+1
dt
n+1
dt
1
=
n!
t
n
.
This is the claimed conditional distribution.
iv) = v) It is clear that II E[N
1
] = . Let x
k
IIN and t
0
= 0 < t
1
< < t
n
.
Then for x = x
1
+ +x
n
II P[N
t
k
N
t
k1
= x
k
, 1 k n]
= II P[N
t
k
N
t
k1
= x
k
, 1 k n [ N
tn
= x]II P[N
tn
= x]
=
(t
n
)
x
x!
e
tn
x!
x
1
! x
n
!
n
k=1
_
t
k
t
k1
t
n
_
x
k
=
n
k=1
((t
k
t
k1
))
x
k
x
k
!
e
(t
k
t
k1
)
and therefore N
t
has independent increments.
v) = i) Note that for x
k
IIN, t
0
= 0 < t
1
< < t
n
and h > 0
II P[N
t
k
+h
N
t
k1
+h
= x
k
, 1 k n [ N
tn+h
]
= II P[N
t
k
N
t
k1
= x
k
, 1 k n [ N
tn+h
] .
Integrating with respect to N
tn+h
yields that N
t
has stationary increments. For
II P[N
t
= 0] we obtain
II P[N
t+s
= 0] = II P[N
t+s
N
s
= 0, N
s
= 0] = II P[N
t
= 0]II P[N
s
= 0] .
It is left as an exercise to show that II P[N
t
= 0] = (II P[N
1
= 0])
t
. Let now
0
:=
log II P[N
1
= 0]. Then II P[N
t
= 0] = e
0
t
. Moreover, for 0 < h < t,
e
0
(th)
= II E[II P[N
th
= 0 [ N
t
]] = e
0
t
+
h
t
II P[N
t
= 1] +
k=2
_
h
t
_
k
II P[N
t
= k] .
Reordering the terms yields
e
0
(th)
e
0
t
h
t = II P[N
t
= 1] +
h
t
k=2
_
h
t
_
k2
II P[N
t
= k] .
Because the sum is bounded by 1, letting h 0 gives II P[N
t
= 1] =
0
te
0
t
. In
particular, N
t
is a Poisson process with rate
0
. From i) = v) we conclude
186 C. RENEWAL PROCESSES
that
0
= II E[N
1
] = . This implies i).
i) = vi) Follows readily from i) and v).
vi) = i) As in the proof of v) = i) it follows that II P[N
t
= 0] = e
0
t
. Then
II P[N
h
= 1] = 1 II P[N
h
= 0] II P[N
h
2] =
0
h + o(h). Thus N
t
is a Poisson
process with rate
0
. From i) v) we conclude that
0
= II E[N
1
] = .
Remark. The condition II E[N
1
] = in v) and vi) is only used for identifying the
parameter . We did not use in the proof that II E[N
1
] < .
The Poisson process has the following properties.
Proposition C.3. Let N
t
and
N
t
be two independent Poisson processes with
rates and
respectively. Let I
i
: i IIN be an iid. sequence of random variables
independent of N
t
with II P[I
i
= 1] = 1 II P[I
i
= 0] = q for some q (0, 1).
Furthermore let a > 0 be a real number. Then
i) N
t
+
N
t
is a Poisson process with rate +
.
ii)
_
Nt
i=1
I
i
_
is a Poisson process with rate q.
iii) N
at
is a Poisson process with rate a.
Proof. Exercise.
Denition C.4. Let (t) be an increasing right continuous function on [0, )
with (0) = 0. A point process N
t
on [0, ) is called inhomogeneous Poisson
process with intensity measure (t) if
i) N
t
has independent increments,
ii) N
t
N
s
Pois((t) (s)).
If there exists a function (t) such that (t) =
_
t
0
(s) ds then (t) is called inten-
sity or rate of the inhomogeneous Poisson process.
Note that a homogeneous Poisson process is a special case with (t) = t. Dene
1
(x) = supt 0 : (t) x the inverse function of (t).
C. RENEWAL PROCESSES 187
Proposition C.5. Let
N
t
be a homogeneous Poisson process with rate 1. Dene
N
t
=
N
(t)
. Then N
t
is an inhomogeneous Poisson process with intensity mea-
sure (t). Conversely, let N
t
be an inhomogeneous Poisson process with intensity
measure (t). Let
N
t
= N
1
(t)
at all points where (
1
(t)) = t. On inter-
vals ((
1
(t)), (
1
(t))) where (
1
(t)) ,= t let there be N
1
(t)
N
(
1
(t))
occurrence points uniformly distributed on the interval ((
1
(t)), (
1
(t))) in-
dependent of N
t
. Then
N
t
is a homogeneous Poisson process with rate 1.
Proof. Exercise.
For an inhomogeneous Poisson process we can construct the following martin-
gales.
Lemma C.6. Let r IR. The following processes are martingales.
i) N
t
(t),
ii) (N
t
(t))
2
(t),
iii) exprN
t
(t)(e
r
1).
Proof. i) Because N
t
has independent increments
II E[N
t
(t) [ T
s
] = II E[N
t
N
s
] +N
s
(t) = N
s
(s) .
ii) Analogously,
II E[(N
t
(t))
2
(t) [ T
s
]
= II E[(N
t
N
s
(t) (s) +N
s
(s))
2
[ T
s
] (t)
= II E[(N
t
N
s
(t) (s))
2
]
+ 2(N
s
(s))II E[N
t
N
s
(t) (s)] + (N
s
(s))
2
(t)
= (t) (s) + 2 0 + (N
s
(s))
2
(t) = (N
s
(s))
2
(s) .
iii) Analogously
II E[e
rNt(t)(e
r
1)
[ T
s
] = II E[e
r(NtNs)
]e
rNs(t)(e
r
1)
= e
rNs(s)(e
r
1)
.
n=0
F
n
(t) .
Moreover, U(t) < for all t 0 and U(t) as t .
Proof. Exercise.
In the renewal theory one often has to solve equations of the form
Z(x) = z(x) +
_
x
0
Z(x y) dF(y) (x 0) (C.2)
where z(x) is a known function and Z(x) is unknown. This equation is called the
renewal equation. The equation can be solved explicitly.
Proposition C.9. If z(x) is bounded on bounded intervals then
Z(x) =
_
x
0
z(x y) dU(y) = z U(x)
is the unique solution to (C.2) that is bounded on bounded intervals.
C. RENEWAL PROCESSES 189
Proof. We rst show that Z(x) is a solution. This follows from
Z(x) = z U(x) = z
n=0
F
n
(x) =
n=0
z F
n
(x)
= z(x) +
n=1
z F
(n1)
F(x) = z(x) +z U F(x) = z(x) +Z F(x) .
Let Z
1
(x) be a solution that is bounded on bounded intervals. Then
[Z
1
(x)Z(x)[ =
_
x
0
(Z
1
(xy)Z(xy)) dF(y)
_
x
0
[Z
1
(xy)Z(xy)[ dF(y)
and by induction
[Z
1
(x) Z(x)[
_
x
0
[Z
1
(x y) Z(x y)[ dF
n
(y) sup
0yx
[Z
1
(y) Z(y)[F
n
(x) .
The latter tends to 0 as n . Thus Z
1
(x) = Z(x).
Let z(x) be a bounded function and h > 0 be a real number. Dene
m
k
(h) = supz(t) : (k 1)h t < kh, m
k
(h) = infz(t) : (k 1)h t < kh
and the Riemann sums
(h) = h
k=1
m
k
(h), (h) = h
k=1
m
k
(h).
Denition C.10. A function z(x) is called directly Riemann integrable if
< lim
h0
(h) = lim
h0
(h) < .
The following lemma gives some criteria for a function to be directly Riemann inte-
grable.
Lemma C.11.
The space of directly Riemann integrable functions is a linear space.
If z(t) is monotone and
_
0
z(t) dt < then z(t) is directly Riemann integrable.
If a(t) and b(t) are directly Riemann integrable and z(t) is continuous Lebesgue
a.e. such that a(t) z(t) b(t) then z(t) is directly Riemann integrable.
If z(t) 0, z(t) is continuous Lebesgue a.e. and (h) < for some h > 0 then
is z(t) directly Riemann integrable.
Proof. See for instance [1, p.69].
190 C. RENEWAL PROCESSES
Denition C.12. A distribution function F of a positive random variable X is
called arithmetic if for some one has II P[X , 2, . . .] = 1. The span is
the largest number such that the above relation is fullled.
The most important result in renewal theory is a result on the asymptotic behaviour
of the solution to (C.2).
Proposition C.13. (Renewal theorem) If z(x) is a directly Riemann inte-
grable function then the solution Z(x) to the renewal equation (C.2) satises
lim
t
Z(t) =
_
0
z(y) dy
if F is non-arithmetic and
lim
n
Z(t +n) =
j=0
z(t +j)
for 0 t < if F is arithmetic with span .
Proof. See for instance [40, p.364] or [66, p.218].
Example C.14. Assume that F is non-arithmetic and that II E[T
2
1
] < . We
consider the function Z(t) = II E[T
Nt+1
t], the expected time till the next occurrence
of an event. Consider rst II E[T
Nt+1
t [ T
1
= s]. If s t then there is a renewal
process starting at s and thus II E[T
Nt+1
t [ T
1
= s] = Z(t s). If s > t then
T
Nt+1
= T
1
= s and thus II E[T
Nt+1
t [ T
1
= s] = s t. Hence we get the renewal
equation
Z(t) = II E[II E[T
Nt+1
t [ T
1
]] =
_
t
(s t) dF(s) +
_
t
0
Z(t s) dF(s) .
Now
z(t) =
_
t
(s t) dF(s) =
_
t
_
s
t
dy dF(s) =
_
t
(1 F(y)) dy .
The function is monotone and
_
0
_
t
(s t) dF(s) dt =
_
0
_
s
0
(s t) dt dF(s) =
1
2
II E[T
2
1
] .
Thus z(t) is directly Riemann integrable and by the renewal theorem
lim
t
Z(t) =
2
II E
_
T
2
1
_
Z(t s) if s t,
0 if t < s t +x,
1 if t +x < s.
This gives the renewal equation
Z(t) = 1 F(t +x) +
_
t
0
Z(t s) dF(s) .
z(t) = 1 F(t +x) is decreasing in t and its integral is bounded by
1
. Thus it is
directly Riemann integrable. It follows from the renewal theorem that
lim
t
Z(t) =
_
0
(1 F(t +x)) dt =
_
x
(1 F(t)) dt .
Thus the stationary distribution of the next event must be the distribution given by
(C.1).
C.3. Bibliographical Remarks
The theory of point processes can also be found in [19] or [26]. For further literature
on renewal theory see also [1] and [40].
192 D. BROWNIAN MOTION
D. Brownian Motion
Denition D.1. A stochastic process W
t
is called a (m,
2
)-Brownian motion
if a.s.
W
0
= 0,
W
t
has independent increments,
W
t
N(mt,
2
t) and
W
t
has cadlag paths.
A (0, 1)-Brownian motion is called a standard Brownian motion.
It can be shown that Brownian motion exists. Moreover, one can proof that a
Brownian motion has continuous paths.
From the denition it follows that W
t
has also stationary increments.
Lemma D.2. A (m,
2
)-Brownian motion has stationary increments.
Proof. Exercise.
We construct the following martingales.
Lemma D.3. Let W
t
be a (m,
2
)-Brownian motion and r IR. The following
processes are martingales.
i) W
t
mt.
ii) (W
t
mt)
2
2
t.
iii) expr(W
t
mt)
2
2
r
2
t.
Proof. i) Using the stationary and independent increments property
II E[W
t
mt [ T
s
] = II E[W
t
W
s
] +W
s
mt = W
s
ms .
ii) It suces to consider the case m = 0.
II E[W
2
t
2
t [ T
s
] = II E[(W
t
W
s
)
2
+ 2(W
t
W
s
)W
s
+W
2
s
[ T
s
]
2
t = W
2
s
2
s .
iii) It suces to consider the case m = 0.
II E[e
rWt
2
2
r
2
t
[ T
s
] = II E[e
r(WtWs)
]e
rWs
2
2
r
2
t
= e
rWs
2
2
r
2
s
.
i=1
X
i
and S
0
= 0.
Lemma E.1.
i) If II E[X
1
] < 0 then S
n
converges to a.s..
ii) If II E[X
1
] = 0 then a.s.
lim
n
S
n
= lim
n
S
n
= .
iii) If II E[X
1
] > 0 then S
n
converges to a.s. and there is a strictly positive proba-
bility that S
n
0 for all n IIN.
Proof. See for instance [40, p.396] or [66, p.233].
For the rest of this section we assume < II E[X
i
] < 0. Let
+
= infn >
0 : S
n
> 0 and
= infn > 0 : S
n
0, H(x) = II P[
+
< , S
+
x] and
(x) = II P[S
n
(x) = II P[S
1
> 0, S
2
> 0, . . . , S
n
> 0, S
n
x]
for n 1. Let
(x) =
n=0
n
(x) .
Lemma E.2. We have for n 1
n
(x) = II P[S
n
> S
j
, (0 j n 1), S
n
x]
and therefore
(x) =
n=0
H
n
(x) .
Moreover, () = (1 H())
1
, II E[
] = () and II E[S
] = II E[
]II E[X
i
].
194 E. RANDOM WALKS AND THE WIENER-HOPF FACTORIZATION
Proof. Let for n xed
S
k
= S
n
S
nk
=
n
i=nk+1
X
i
.
Then S
k
: k n follows the same law as S
k
: k n. Thus
II P[S
n
> S
j
, (0 j n 1), S
n
x] = II P[S
n
> S
j
, (0 j n 1), S
n
x]
= II P[S
n
> S
n
S
nj
, (0 j n 1), S
n
x]
= II P[S
j
> 0, (1 j n), S
n
x] =
n
(x) .
Denote by
n+1
:= infk > 0 : S
k
> S
n
the n + 1-st ascending ladder time. Note
that S
n
has distribution H
n
(x). We found that
n
(x) is the probability that S
n
is a maximum of the random walk and lies in the interval (0, x]. Thus
n
(x) is the
probability that there is a ladder height at n and S
n
x. We obtain
n=1
n
(x) =
n=1
II P[k :
k
= n, 0 < S
n
x] =
n=1
k=1
II P[
k
= n, 0 < S
k
x]
=
k=1
II P[0 < S
k
x,
k
< ] =
k=1
H
k
(x) .
Because II E[X
i
] < 0 we have H() < 1 (Lemma E.1) and H
n
() = (H())
n
from
which () = (1 H())
1
follows.
For the expected value of
II E[
] =
n=0
II P[
> n] =
n=0
n
() = () .
Finally let
(n1)
S
(n)
. Then
L
1
+ +L
n
n
=
S
(n)
(n)
(n)
n
from which the last assertion follows by the strong law of large numbers.
Let now
n
(x) = II P[S
1
> 0, S
2
> 0, . . . , S
n1
> 0, S
n
x] .
Note that for x 0
(x) =
n=1
n
(x) ,
E. RANDOM WALKS AND THE WIENER-HOPF FACTORIZATION 195
that
1
(x) = U(x) U(0) if x > 0 and
1
(x) = U(x) if x 0. For n 1 one obtains
for x > 0
n+1
(x) =
_
0
(U(x y) U(y))
n
(dy)
and for x 0
n+1
(x) =
_
0
U(x y)
n
(dy) .
If we sum over all n we obtain for x > 0
(x) 1 =
_
0
(U(x y) U(y))(dy)
and for x 0
(x) =
_
0
U(x y)(dy) .
Note that (0) = (0) = 1 and thus for x 0
(x) =
_
0
U(x y)(dy) .
The last two equations can be written as a single equation
+ =
0
+ U .
Then
0
+ H = H + H = H + H U = H + U U
= H U + +
0
from which it follows that
U = H + H . (E.1)
The latter is called the Wiener-Hopf factorization.
E.1. Bibliographical Remarks
The results of this section are taken from [40].
196 F. SUBEXPONENTIAL DISTRIBUTIONS
F. Subexponential Distributions
Denition F.1. A distribution function F with F(x) = 0 for x < 0 is called
subexponential if
lim
t
1 F
2
(t)
1 F(t)
= 2 .
The intuition behind the denition is the following. One often observes that the
aggregate sum of claims is determined by a few of the largest claims. The denition
can be rewritten as II P[X
1
+ X
2
> x] II P[maxX
1
, X
2
> x]. We will prove in
Lemma F.7 below that subexponential is equivalent to II P[X
1
+X
2
+ +X
n
> x]
II P[maxX
1
, X
2
, . . . , X
n
> x] for some n 2.
We want to show that the moment generating function of such a distribution
does not exist for strictly positive values. We rst need the following
Lemma F.2. If F is subexponential then for all t IR
lim
x
1 F(x t)
1 F(x)
= 1 .
Proof. Let t 0. We have
1 F
2
(x)
1 F(x)
1 =
F(x) F
2
(x)
1 F(x)
=
_
t
0
1 F(x y)
1 F(x)
dF(y) +
_
x
t
1 F(x y)
1 F(x)
dF(y)
F(t) +
1 F(x t)
1 F(x)
(F(x) F(t)) .
Thus
1
1 F(x t)
1 F(x)
(F(x) F(t))
1
_
1 F
2
(x)
1 F(x)
1 F(t)
_
.
The assertion for t 0 follows by letting x . If t < 0 then
lim
x
1 F(x t)
1 F(x)
= lim
x
1
1 F((x t) (t))
1 F(x t)
= lim
y
1
1 F(y (t))
1 F(y)
= 1
where y = x t.
F. SUBEXPONENTIAL DISTRIBUTIONS 197
Lemma F.3. Let F be subexponential and r > 0. Then
lim
t
e
rt
(1 F(t)) = ,
in particular
_
0
e
rx
dF(x) = .
Proof. We rst observe that for any 0 < x < t
e
rt
(1 F(t)) =
1 F(t)
1 F(t x)
(1 F(t x))e
r(tx)
e
rx
(F.1)
and by Lemma F.2 the function e
rn
(1 F(n)) (n IIN) is increasing for n large
enough. Thus there exist a limit in (0, ]. Letting n in (F.1) shows that this
limit can only be innite.
Let now t
n
be an arbitrary sequence such that t
n
. Then
e
rtn
(1 F(t
n
)) =
1 F(t
n
)
1 F([t
n
])
(1 F([t
n
]))e
r[tn]
e
r(tn[tn])
1 F(t
n
)
1 F(t
n
1)
(1 F([t
n
]))
which tends to innity as n .
The moment generating function can be written as
_
0
e
rx
dF(x) = 1 +
_
0
_
x
0
re
ry
dy dF(x)
= 1 +r
_
0
_
y
dF(x)e
ry
dy = 1 +r
_
0
e
ry
(1 F(y)) dy =
because the integrand diverges to innity.
The following lemma gives a condition for subexponentiality.
Lemma F.4. Assume that for all z (0, 1] the limit
(z) = lim
x
1 F(zx)
1 F(x)
exists and that (z) is left-continuous at 1. Then F is a subexponential distribution
function.
198 F. SUBEXPONENTIAL DISTRIBUTIONS
Proof. Note rst that F
2
(x) = II P[X
1
+ X
2
x] II P[X
1
x, X
2
x] = F
2
(x).
For simplicity assume F(0) = 0. Hence
lim
x
1 F
2
(x)
1 F(x)
lim
x
1 F
2
(x)
1 F(x)
= lim
x
1 +F(x) = 2 .
Let n 1 be xed. Then
lim
x
1 F
2
(x)
1 F(x)
= 1 + lim
x
_
x
0
1 F(x y)
1 F(x)
dF(y)
1 + lim
x
n
k=1
1 F(x kx/n)
1 F(x)
(F(kx/n) F((k 1)x/n))
= 1 +(1 1/n) .
Because n is arbitrary and (z) is left continuous at 1 the assertion follows.
Example F.5. Consider the Pa(, ) distribution.
1 F(zx)
1 F(x)
=
(
+zx
)
(
+x
)
=
_
+x
+zx
_
n
= sup
t0
1 F
n
(t)
1 F(t)
and note that 1 F
(n+1)
(t) = 1 F(t) +F (1 F
n
)(t). Choose T 0 such that
sup
tT
F(t) F
2
(t)
1 F(t)
< 1 +
2
.
Let A
T
= (1 F(T))
1
. For n 1
n+1
1 + sup
0tT
_
t
0
1 F
n
(t y)
1 F(t)
dF(y) + sup
tT
_
t
0
1 F
n
(t y)
1 F(t)
dF(y)
1 +A
T
+ sup
tT
_
t
0
1 F
n
(t y)
1 F(t y)
1 F(t y)
1 F(t)
dF(y)
1 +A
T
+
n
sup
tT
F(t) F
2
(t)
1 F(t)
1 +A
T
+
n
_
1 +
2
_
.
F. SUBEXPONENTIAL DISTRIBUTIONS 199
Choose
D = max
_
2(1 +A
T
)
, 1
_
and note that
1
= 1 < D(1 +). The assertion follows by induction.
Lemma F.7. Let F(x) = 0 for x < 0. The following are equivalent:
i) F is subexponential.
ii) For all n IIN
lim
x
1 F
n
(x)
1 F(x)
= n. (F.2)
iii) There exists n 2 such that (F.2) holds.
Proof. i) = ii) We proof the assertion by induction. The assertion is trivial
for n = 2. Assume that it is proved for some n IIN with n 2. Let > 0. Choose
t such that
1 F
n
(x)
1 F(x)
n
<
for x t.
1 F
(n+1)
(x)
1 F(x)
= 1 +
_
xt
0
1 F
n
(x y)
1 F(x y)
1 F(x y)
1 F(x)
dF(y)
+
_
x
xt
1 F
n
(x y)
1 F(x)
dF(y) .
The second integral is bounded by
_
x
xt
1 F
n
(x y)
1 F(x)
dF(y)
F(x) F(x t)
1 F(x)
=
1 F(x t)
1 F(x)
1
which tends to 0 as x by Lemma F.2. The expression
_
xt
0
n
1 F(x y)
1 F(x)
dF(y) = n
_
F(x) F
2
(x)
1 F(x)
_
x
xt
1 F(x y)
1 F(x)
dF(y)
_
n
as x by the same arguments as above. Then
_
xt
0
_
1 F
n
(x y)
1 F(x y)
n
_
1 F(x y)
1 F(x)
dF(y)
_
F(x) F
2
(x)
1 F(x)
_
x
xt
1 F(x y)
1 F(x)
dF(y)
_
200 F. SUBEXPONENTIAL DISTRIBUTIONS
by the arguments used before. Thus
lim
x
1 F
(n+1)
(x)
1 F(x)
(n + 1)
.
This proves the assertion because was arbitrary.
ii) = iii) Trivial.
iii) = i) Assume now (F.2) for some n > 2. We show that the assertion also
holds for n replaced by n 1. Since
1 F
n
(x)
1 F(x)
1 =
_
x
0
1 F
(n1)
(x y)
1 F(x)
dF(y) F(x)
1 F
(n1)
(x)
1 F(x)
it follows that
lim
x
1 F
(n1)
(x)
1 F(x)
n 1 .
Because F
(n1)
(x) F(x)
n1
it is trivial that
lim
x
1 F
(n1)
(x)
1 F(x)
n 1 .
It follows that (F.2) holds for n = 2 and F is a subexponential distribution.
Lemma F.8. Let U and V be two distribution functions with U(x) = V (x) = 0
for all x < 0. Assume that
1 V (x) a(1 U(x)) as x
for some a > 0. If U is subexponential then also V is subexponential.
Proof. We have to show that
lim
x
_
x
0
1 V (x y)
1 V (x)
dV (y) 1 .
Let 0 < < 1. There exists y
0
such that for all y y
0
(1 )a
1 V (y)
1 U(y)
(1 +)a .
Let x > y
0
. Note that
_
x
xy
0
1 V (x y)
1 V (x)
dV (y)
V (x) V (x y
0
)
1 V (x)
=
1V (xy
0
)
1U(xy
0
)
1V (x)
1U(x)
1 U(x y
0
)
1 U(x)
1 0
F. SUBEXPONENTIAL DISTRIBUTIONS 201
by Lemma F.2. Moreover,
_
xy
0
0
1 V (x y)
1 V (x)
dV (y)
1 +
1
_
xy
0
0
1 U(x y)
1 U(x)
dV (y)
1 +
1
_
x
0
1 U(x y)
1 U(x)
dV (y) .
The last integral is
_
x
0
1 U(x y)
1 U(x)
dV (y) =
V (x) U V (x)
1 U(x)
= 1
1 V (x)
1 U(x)
+
U(x) V U(x)
1 U(x)
= 1
1 V (x)
1 U(x)
+
_
x
0
1 V (x y)
1 U(x)
dU(y) .
1 (1 U(x))
1
(1 V (x)) tends to 1 a as x . And
_
x
xy
0
1 V (x y)
1 U(x)
dU(y)
U(x) U(x y
0
)
1 U(x)
tends to 0. Finally
_
xy
0
0
1 V (x y)
1 U(x)
dU(y) (1 +)a
_
xy
0
0
1 U(x y)
1 U(x)
dU(y)
(1 +)a
_
x
0
1 U(x y)
1 U(x)
dU(y)
which tends to (1 +)a. Putting all these limits together we obtain
lim
x
_
x
0
1 V (x y)
1 V (x)
dV (y)
(1 +)(1 +a)
1
.
Because was arbitrary the assertion follows.
F.1. Bibliographical Remarks
The class of subexponential distributions was introduced by Chistyakov [23]. The
results presented here can be found in [23], [13] and [34].
202 G. CONCAVE AND CONVEX FUNCTIONS
G. Concave and Convex Functions
In this appendix we let I be an interval, nite or innite, but not a singleton.
Denition G.1. A function u : I IR is called (strictly) concave if for all
x, z I, x ,= z, and all (0, 1) one has
u((1 )x +z) (>) (1 )u(x) +u(z) .
u is called (strictly) convex if u is (strictly) concave.
Because results on concave functions can easily translated for convex functions we
will only consider concave functions in the sequel.
Concave functions have nice properties.
Lemma G.2. A concave function u(y) is continuous, dierentiable from the left
and from the right. The derivative is decreasing, i.e. for x < y we have u
t
(x)
u
t
(x+) u
t
(y) u
t
(y+). If u(y) is strictly concave then u
t
(x+) > u
t
(y).
Remark. The theorem implies that u(y) is dierentiable almost everywhere.
Proof. Let x < y < z. Then
u(y) = u
_
z y
z x
x +
y x
z x
z
_
z y
z x
u(x) +
y x
z x
u(z)
or equivalently
(z x)u(y) (z y)u(x) + (y x)u(z) . (G.1)
This implies immediately
u(y) u(x)
y x
u(z) u(x)
z x
u(z) u(y)
z y
. (G.2)
Thus the function h h
1
(u(y) u(y h)) is increasing in h and bounded from
below by (z y)
1
(u(z) u(y)). Thus the derivative u
t
(y) from the left exists.
Analogously, the derivative from the right u
t
(y+) exists. The assertion in the concave
case follows now from (G.2). The strict inequality in the strictly concave case follows
analogously.
Concave functions have also the following property.
G. CONCAVE AND CONVEX FUNCTIONS 203
Lemma G.3. Let u(y) be a concave function. There exists a function k : I IR
such that for any y, x I
u(x) u(y) +k(y)(x y) . (G.3)
Moreover, the function k(y) is decreasing. If u(y) is strictly concave then the above
inequality is strict for x ,= y and k(y) is strictly decreasing. Conversely, if a function
k(y) exists such that (G.3) is fullled, then u(y) is concave, strictly concave if the
strict inequality holds for x ,= y.
Proof. Left as an exercise.
Corollary G.4. Let u(y) be a twice dierentiable function. Then u(y) is concave
if and only if its second derivative is negative. It is strictly concave if and only if its
second derivative is strictly negative almost everywhere.
Proof. This follows readily from Theorem G.2 and Lemma G.3.
The following result is very useful.
Theorem G.5. (Jensens inequality) The function u(y) is (strictly) concave if
and only if
II E[u(Y )] (<) u(II E[Y ]) (G.4)
for all I-valued integrable random variables Y with II P[Y ,= II E[Y ]] > 0.
Proof. Assume (G.4) for all random variables Y . Let (0, 1). Let II P[Y = z] =
1 II P[Y = x] = . Then the (strict) concavity follows. Assume u(y) is strictly
concave. Then it follows from Lemma G.3 that
u(Y ) u(II E[Y ]) +k(II E[Y ])(Y II E[Y ]) .
The strict inequality holds if u(y) is strictly concave and Y ,= II E[Y ]. Taking expected
values gives (G.4).
Also the following result is often useful.
204 G. CONCAVE AND CONVEX FUNCTIONS
Theorem G.6. (Ohlins lemma) Let F
i
(y), i = 1, 2 be two distribution func-
tions dened on I. Assume
_
I
y dF
1
(y) =
_
I
y dF
2
(y) <
and that there exists y
0
I such that
F
1
(y) F
2
(y), y < y
0
, F
1
(y) F
2
(y), y > y
0
.
Then for any concave function u(y)
_
I
u(y) dF
1
(y)
_
I
u(y) dF
2
(y)
provided the integrals are well dened. If u(y) is strictly concave and F
1
,= F
2
then
the inequality holds strictly.
Proof. Recall the formulae
_
0
y dF
i
(y) =
_
0
(1 F
i
(y)) dy and
_
0
y dF
i
(y) =
_
0
F
i
(y) dy, which can be proved for example by Fubinis theorem. Thus it
follows that
_
I
(F
2
(y) F
1
(y)) dy = 0. We know that u(y) is dierentiable almost
everywhere and continuous. Thus u(y) = u(y
0
) +
_
y
y
0
u
t
(z) dz, where we can dene
u
t
(y) as the right derivative. This yields
_
y
0
u(y) dF
i
(y) = u(y
0
)F
i
(y
0
)
_
y
0
_
y
0
y
u
t
(z) dz dF
i
(y)
= u(y
0
)F
i
(y
0
)
_
y
0
F
i
(z)u
t
(z) dz .
Analogously
_
y
0
u(y) dF
i
(y) = u(y
0
)(1 F
i
(y
0
)) +
_
y
0
(1 F
i
(z))u
t
(z) dz .
Putting the results together we nd
_
u(y) dF
1
(y)
_
u(y) dF
2
(y) =
_
(F
2
(y) F
1
(y))u
t
(y) dy .
If y < y
0
then F
2
(y)F
1
(y) 0 and u
t
(y) u
t
(y
0
). If y > y
0
then F
2
(y)F
1
(y) 0
and u
t
(y) u
t
(y
0
). Thus
_
u(y) dF
1
(y)
_
u(y) dF
2
(y)
_
(F
2
(y) F
1
(y))u
t
(y
0
) dy = 0 .
The strictly concave case follows analogously.
G. CONCAVE AND CONVEX FUNCTIONS 205
Corollary G.7. Let X be a real random variable taking values in some interval
I
1
, and let g
i
: I
1
I
2
, i = 1, 2 be increasing functions with values on some interval
I
2
. Suppose
II E[g
1
(X)] = II E[g
2
(X)] < .
Let u : I
2
IR be a concave function such that II E[u(g
i
(X))] is well-dened. If there
exists x
0
such that
g
1
(x) g
2
(x), x < x
0
, g
1
(x) g
2
(x), x > x
0
,
then
II E[u(g
1
(X))] II E[u(g
2
(X))] .
Moreover, if u(y) is strictly concave and II P[g
1
(X) ,= g
2
(X)] > 0 then the inequality
is strict.
Proof. Choose F
i
(y) = II P[g
i
(X) y]. Let y
0
= g
1
(x
0
). If y < y
0
then
F
1
(y) = II P[g
1
(X) y] = II P[g
1
(X) y, X < x
0
] II P[g
2
(X) y, X < x
0
] F
2
(y) .
If y > y
0
then
1 F
1
(y) = II P[g
1
(X) > y, X > x
0
] II P[g
2
(X) > y, X > x
0
] 1 F
2
(y) .
The result follows now from Theorem G.6.
206 TABLE OF DISTRIBUTION FUNCTIONS
Table of Distribution Functions
Exponential distribution: Exp()
Parameters: > 0
Distribution function: F(x) = 1 e
x
Density : f(x) = e
x
Expected value: II E[X] =
1
Variance: Var(X) =
2
Moment generating function: M
X
(r) = ( r)
1
Pareto distribution: Pa(, )
Parameters: > 0, > 0
Distribution function: F(x) = 1
( +x)
Density : f(x) =
( +x)
1
Expected value: II E[X] = ( 1)
1
if > 1
Variance: Var(X) =
2
( 1)
2
( 2)
1
if > 2
Moment generating function: M
X
(r) =
Weibull distribution: Wei(, c)
Parameters: > 0, c > 0
Distribution function: F(x) = 1 e
cx
Density : f(x) = cx
1
e
cx
)
Variance: Var(X) = c
2/
((1 +
2
) (1 +
1
)
2
)
Moment generating function: M
X
(r) =
Gamma distribution: (, )
Parameters: > 0, > 0
Distribution function: F(x) =
Density : f(x) =
(())
1
x
1
e
x
Expected value: II E[X] =
1
Variance: Var(X) =
2
Moment generating function: M
X
(r) = (( r)
1
)
(())
1
x
1
(log x)
1
(x > 1)
Expected value: II E[X] = (( 1)
1
)
if > 1
Variance: Var(X) = (/( 2))
(/( 1))
2
, > 2
Moment generating function: M
X
(r) =
TABLE OF DISTRIBUTION FUNCTIONS 207
Normal distribution: N(,
2
)
Parameters: IR,
2
> 0
Distribution function: F(x) = (
1
(x ))
Density : f(x) = (2
2
)
1/2
e
(x)
2
/(2
2
)
Expected value: II E[X] =
Variance: Var(X) =
2
Moment generating function: M
X
(r) = exp(
2
/2)r
2
+r
Lognormal distribution: LN(,
2
)
Parameters: IR,
2
> 0
Distribution function: F(x) = (
1
(log x ))
Density : f(x) = (2
2
)
1/2
x
1
e
(log x)
2
/(2
2
)
Expected value: II E[X] = e
(
2
/2)+
Variance: Var(X) = e
2
+2
(e
2
1)
Moment generating function: M
X
(r) =
Binomial distribution: B(k, p)
Parameters: k IIN, p (0, 1)
Probabilities: II P[X = n] =
_
k
n
_
p
n
(1 p)
kn
Expected value: II E[X] = kp
Variance: Var(X) = kp(1 p)
Moment generating function: M
X
(r) = (pe
r
+ 1 p)
k
Poisson distribution: Pois()
Parameters: > 0
Probabilities: II P[X = n] = (n!)
1
n
e
(1 p)
n
Expected value: II E[X] = (1 p)/p
Variance: Var(X) = (1 p)/p
2
Moment generating function: M
X
(r) = p
(1 (1 p)e
r
)
208 REFERENCES
References
[1] Alsmeyer, G. (1991). Erneuerungstheory. Teubner, Stuttgart.
[2] Ammeter, H. (1948). A generalization of the collective theory of risk in regard
to uctuating basic probabilities. Skand. Aktuar Tidskr., 171198.
[3] Andersen, E.Sparre (1957). On the collective theory of risk in the case of
contagion between the claims. Transactions XVth International Congress of
Actuaries, New York, II, 219229.
[4] Arfwedson, G. (1954). Research in collective risk theory. Part 1. Skand.
Aktuar Tidskr., 191223.
[5] Arfwedson, G. (1955). Research in collective risk theory. Part 2. Skand.
Aktuar Tidskr., 53100.
[6] Asmussen, S. (1982). Conditioned limit theorems relating a random walk to
its associate, with applications to risk reserve processes and the GI/G/1 Queue.
Adv. in Appl. Probab. 14, 143170.
[7] Asmussen, S. (1984). Approximations for the probability of ruin within nite
time. Scand. Actuarial J., 3157.
[8] Asmussen, S. (1989). Risk theory in a Markovian environment. Scand. Ac-
tuarial J., 66100.
[9] Asmussen, S. (2000). Ruin Probabilities. World Scientic, Singapore.
[10] Asmussen, S., Fle Henriksen, F. and Kl uppelberg, C. (1994). Large
claims approximations for risk processes in a Markovian environment. Stochas-
tic Process. Appl. 54, 2943.
[11] Asmussen, S. and Nielsen, H.M. (1995). Ruin probabilities via local ad-
justment coecients. J. Appl. Probab., 736755.
[12] Asmussen, S., Schmidli, H. and Schmidt, V. (1999). Tail probabilities
for non-standard risk and queueing processes with subexponential jumps. Adv.
in Appl. Probab. 31, to appear.
[13] Athreya, K.B. and Ney, P. (1972). Branching Processes. Springer-Verlag,
Berlin.
[14] Bailey, A.L. (1945). A generalised theory of credibility. Proceedings of the
Casualty Actuarial Society 32, 1320.
[15] Bailey, A.L. (1950). Credibility procedures, La Places generalisation of
Bayes rule, and the combination of collateral knowledge with observed data.
Proceedings of the Casualty Actuarial Society 37, 723 and 94115.
[16] Barndor-Nielsen, O.E. and Schmidli, H. (1994). Saddlepoint approxi-
mations for the probability of ruin in nite time. Scand. Actuarial J., 169186.
REFERENCES 209
[17] Beekman, J. (1969). A ruin function approximation. Trans. Soc. Actuar-
ies 21, ,. 4148 and 275279
[18] Bjork, T. and Grandell, J. (1988). Exponential inequalities for ruin prob-
abilities in the Cox case. Scand. Actuarial J., 77111.
[19] Bremaud, P. (1981). Point Processes and Queues. Springer-Verlag, New York.
[20] B uhlmann, H. (1967). Experience rating and credibility. ASTIN Bulletin 4,
199207.
[21] B uhlmann, H. and Straub, E. (1970). Glaubw urdigkeit f ur Schadensatze.
Schweiz. Verein. Versicherungsmath. Mitt. 70, 111133.
[22] Canteno, L. (1986). Measuring the eects of reinsurance by the adjustment
coecient. Insurance Math. Econom. 5, 169182.
[23] Chistyakov, V.P. (1964). A theorem on sums of independent, positive random
variables and its applications to branching processes. Theory Probab. Appl. 9,
640648.
[24] Cramer, H. (1930). On the Mathematical Theory of Risk. Skandia Jubilee
Volume, Stockholm.
[25] Cramer, H. (1955). Collective Risk Theory. Skandia Jubilee Volume, Stock-
holm.
[26] Daley, D.J. and Vere-Jones, D. (1988). An Introduction to the Theory of
Point Processes. Springer-Verlag, New York.
[27] Dassios, A. (1987). Ph.D. Thesis. Imperial College, London.
[28] Dassios, A. and Embrechts, P. (1989). Martingales and insurance risk.
Commun. Statist. Stochastic Models 5, 181217.
[29] Deheuvels, P., Haeusler, E. and Mason, D.M. (1988). Almost sure con-
vergence of the Hill estimator. Math. Proc. Camb. Phil. Soc. 104, 371381.
[30] Dellacherie, C. and Meyer, P.A. (1980). Probabilites et Potentiel. Ch. VI,
Hermann, Paris.
[31] DeVylder, F. (1978). A practical solution to the problem of ultimate ruin
probability. Scand. Actuarial J., 114119.
[32] Dickson, D.C.M. (1992). On the distribution of the surplus prior to ruin.
Insurance Math. Econom. 11, 191207.
[33] Dufresne, F. and Gerber, H.U. (1988). The surpluses immediately be-
fore and at ruin, and the amount of the claim causing ruin. Insurance Math.
Econom. 7, 193199.
[34] Embrechts, P., Goldie, C.M. and Veraverbeke, N. (1979). Subexponen-
tiality and innite divisibility. Z. Wahrsch. verw. Geb. 49, 335347.
210 REFERENCES
[35] Embrechts, P., Grandell, J. and Schmidli, H. (1993). Finite-time Lund-
berg inequalities in the Cox case. Scand. Actuarial J., 1741.
[36] Embrechts, P., Kl uppelberg, C. and Mikosch, T. (1997). Modelling
Extremal Events. Springer-Verlag, Berlin.
[37] Embrechts, P. and Schmidli, H. (1994). Modelling of extremal events in
insurance and nance. ZORMath. Methods Oper. Res. 39, 133.
[38] Embrechts, P. and Veraverbeke, N. (1982). Estimates for the probability
of ruin with special emphasis on the possibility of large claims. Insurance Math.
Econom. 1, 5572.
[39] Ethier, S.N. and Kurtz, T.G. (1986). Markov Processes. Wiley, New York.
[40] Feller, W. (1971). An Introduction to Probability Theory and its Applica-
tions. Volume II, Wiley, New York.
[41] Gerber, H.U. (1973). Martingales in risk theory. Schweiz. Verein. Ver-
sicherungsmath. Mitt. 73, 205216.
[42] Gerber, H.U. (1975). The surplus process as a fair game utilitywise. ASTIN
Bulletin 8, 307322.
[43] Gerber, H.U. (1979). An Introduction to Mathematical Risk Theory. Hueb-
ner Foundation Monographs, Philadelphia.
[44] Gerber, H.U., Goovaerts, M.J. and Kaas, R. (1987). On the probability
and severity of ruin. ASTIN Bulletin 17, 151163.
[45] Grandell, J. (1977). A class of approximations of ruin probabilities. Scand.
Actuarial J., 3752.
[46] Grandell, J. (1991). Aspects of Risk Theory. Springer-Verlag, New York.
[47] Grandell, J. (1991). Finite time ruin probabilities and martingales. Informat-
ica 2, 332.
[48] Grandell, J. (1995). Some remarks on the Ammeter risk process. Schweiz.
Verein. Versicherungsmath. Mitt. 95, 4371.
[49] Grandell, J. (1997). Mixed Poisson Processes. Chapman & Hall, London.
[50] Gut, A. (1988). Stopped Random Walks. Springer-Verlag, New York.
[51] Hadwiger, H. (1940).
Uber die Wahrscheinlichkeit des Ruins bei einer
grossen Zahl von Geschaften. Archiv f ur mathematische Wirtschafts- und
Sozialforschung 6, 131135.
[52] Hall, P. (1982). On some simple estimates of an exponent of regular variation.
J. R. Statist. Soc. B 44, 3742.
REFERENCES 211
[53] Hill, B.M. (1975). A simple general approach to inference about the tail of a
distribution. Ann. Statist. 3, 11631174.
[54] Hogg, R.V. and Klugman, S.A. (1984). Loss Distributions. Wiley, New
York.
[55] Iglehart, D.L. (1969). Diusion approximations in collective risk theory. J.
Appl. Probab. 6, 285292.
[56] Janssen, J. (1980). Some transient results on the M/SM/1 special semi-
Markov model in risk and queueing theories. ASTIN Bulletin 11, 4151.
[57] Jung, J. (1982). Association of Swedish Insurance Companies. Statistical
Department, Stockholm.
[58] Karlin, S. and Taylor, H.M. (1975). A First Course in Stochastic Processes.
Academic Press, New York.
[59] Karlin, S. and Taylor, H.M. (1981). A Second Course in Stochastic Pro-
cesses. Academic Press, New York.
[60] Kl uppelberg, C. (1988). Subexponential distributions and integrated tails.
J. Appl. Probab. 25, 132141.
[61] Lundberg, F. (1903). I. Approximerad Framstallning av Sannolikhetsfunk-
tionen. II.
Aterforsakring av Kollektivrisker. Almqvist & Wiksell, Uppsala.
[62] Lundberg, F. (1926). Forsakringsteknisk Riskutjamning. F. Englunds bok-
tryckeri A.B., Stockholm.
[63] Norberg, R. (1979). The credibility approach to experience rating. Scand.
Actuarial J.. 181221
[64] Panjer, H.H. (1981). Recursive evaluation of a family of compound distribu-
tions. ASTIN Bulletin 12, 2226.
[65] Panjer, H.H. and Willmot, G.E. (1992). Insurance Risk Models. Society
of Actuaries.
[66] Rolski, T., Schmidli, H., Schmidt, V. and Teugels, J.L. (1999). Sto-
chastic Processes for Insurance and Finance. Wiley, Chichester.
[67] Ross, S.M. (1983). Stochastic Processes. Wiley, New York.
[68] Rytgaard, M. (1996). Simulation experiments on the mean residual life func-
tion m(x). Proceedings of the XXVII ASTIN Colloquium, Copenhagen, Den-
mark, vol. 1, 5981.
[69] Schmidli, H. (1992). A general insurance risk model. Ph.D Thesis, ETH
Z urich.
212 REFERENCES
[70] Schmidli, H. (1994). Diusion approximations for a risk process with the
possibility of borrowing and investment. Comm. Statist. Stochastic Models 10,
365388.
[71] Schmidli, H. (1997). An extension to the renewal theorem and an application
to risk theory. Ann. Appl. Probab. 7, 121133.
[72] Schmidli, H. (1998). On the distribution of the surplus prior and at ruin.
ASTIN Bull., to appear.
[73] Schnieper, R. (1990). Insurance premiums, the insurance market and the
need for reinsurance. Schweiz. Verein. Versicherungsmath. Mitt. 90, 129147.
[74] Seal, H.L. (1972). Numerical calculation of the probability of ruin in the
Poisson/exponential case. Schweiz. Verein. Versicherungsmath. Mitt. 72, 77
100.
[75] Seal, H.L. (1974). The numerical calculation of U(w, t), the probability of
non-ruin in an interval (0, t). Scand. Actuarial J., 121-139.
[76] Segerdahl, C.-O. (1955). When does ruin occur in the collective theory of
risk?. Skand. Aktuar Tidskr., 2236.
[77] Siegmund, D. (1979). Corrected diusion approximation in certain random
walk problems. Adv. in Appl. Probab. 11, 701719.
[78] Sundt, B. (1999). An Introduction to Non-Life Insurance Mathematics. Verlag
Versicherungswirtschaft e.V., Karlsruhe.
[79] Stam, A.J. (1973). Regular variation of the tail of a subordinated probability
distribution. Adv. Appl. Probab. 5, 308327.
[80] Takacs, L. (1962). Introduction to the Theory of Queues. Oxford University
Press, New York.
[81] Thorin, O. (1974). On the asymptotic behavior of the ruin probability for
an innite period when the epochs of claims form a renewal process. Scand.
Actuarial J., 8199.
[82] Thorin, O. (1975). Stationarity aspects of the Sparre Andersen risk process
and the corresponding ruin probabilities. Scand. Actuarial J., 8798.
[83] Thorin, O. (1982). Probabilities of ruin. Scand. Actuarial J., 65102.
[84] Thorin, O. and Wikstad, N. (1977). Calculation of ruin probabilities when
the claim distribution is lognormal. ASTIN Bulletin 9, 231246.
[85] von Bahr, B. (1975). Asymptotic ruin probabilities when exponential mo-
ments do not exist. Scand. Actuarial J., 610.
[86] Waters, H.R. (1983). Some mathematical aspects of reinsurance insurance.
Insurance Math. Econom. 2, 1726.
REFERENCES 213
[87] Whitt, W. (1970). Weak convergence of probability measures on the function
space C[0, ). Ann. Statist. 41, 939944.
[88] Williams, D. (1979). Diusions, Markov Processes and Martingales. Volume
I, Wiley, New York.
214 INDEX
Index
C
t
Ammeter model, 139
Cramer-Lundberg model, 62
Markov modulated model, 168
renewal model, 101
(r), 169
T
T
(T stopping time), 181
(u)
Ammeter model, 140
Cramer-Lundberg model, 62
Markov modulated model, 174
renewal model, 101