See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/252507857

A Lognormal Model for Insurance Claims Data

Article in Revstat - Statistical Journal · January 2006

DOI: 10.57805/revstat.v4i2.31

CITATIONS READS

13 1,454

3 authors, including:

Daiane Aparecida Zuanetti Carlos Diniz

Universidade Federal de São Carlos Universidade Federal de São Carlos
15 PUBLICATIONS 55 CITATIONS 55 PUBLICATIONS 269 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Carlos Diniz on 27 August 2014.

The user has requested enhancement of the downloaded file.

REVSTAT – Statistical Journal
Volume 4, Number 2, June 2006, 131–142

A LOGNORMAL MODEL FOR INSURANCE

CLAIMS DATA

Authors: Daiane Aparecida Zuanetti

– Departamento de Estatı́stica, Universidade Federal de São Carlos,
São Paulo, Brazil
Carlos A. R. Diniz
– Departamento de Estatı́stica, Universidade Federal de São Carlos,
São Paulo, Brazil
[email protected]
José Galvão Leite
– Departamento de Estatı́stica, Universidade Federal de São Carlos,
São Paulo, Brazil

Received: June 2005 Revised: February 2006 Accepted: February 2006

Abstract:
• In the insurance area, especially based on observations of the number of claims, N (w),
corresponding to an exposure w, and on observations of the total amount of claims
incurred, Y (w), the risk theory arises to quantify risks and to fit models of pricing
and insurance company ruin. However, the main problem is the complexity to obtain
the distribution function of Y (w) and, consequently, the likelihood function used to
calculate the estimation of the parameters.
This work considers the Poisson(wλ), λ > 0, for N (w) and lognormal(µ, σ 2 ),
−∞ < µ < ∞, and σ 2 > 0, for Zi , the individual claims, and presents maximum-
likelihood estimates for λ, µ and σ 2 .

Key-Words:

• lognormal distribution; maximum-likelihood estimation; number of claims; total

amount of claims.

AMS Subject Classification:

• 62J02, 62F03.
132 Zuanetti, Diniz and Leite
A Lognormal Model for Insurance Claims Data 133

1. INTRODUCTION

In the insurance area, the main goals of the risk theory are to study, an-
alyze, specify dimensions and quantify risks. The risk theory is also responsible
for fitting models of pricing and insurance company ruin, especially based on
observations of the random variables for the number of claims, N (w), and the
total amount of claims incurred, Y (w), defined as

N (w)
X
(1.1) Y (w) = Zi I(N (w)>0)
i=1

where the Zi ’s are random variables representing the individual claims, w = vt

corresponds to the exposure, v denotes the value insured and t is the period
during which the value v is exposed to the risk of claims.

Assuming that N (w), Z1 , Z2 , ... are independent and the individual claims
are identically distributed, Jorgensen and Souza ([4]) discussed the estimation
and inference problem concerning the parameters considering the situation in
which the number of claims follows a Poisson process and the individual claims
follow a gamma distribution.

Using the properties of the Tweedie family for exponential dispersion mod-
els ([8]; [3]), Jorgensen and Souza ([4]) determined, using the convolution formula,
that Y (w) | N (w) follows an exponential dispersion model and the joint distribu-
tion of N (w) and Y (w)/w follows a Tweedie compound Poisson distribution.
For more details about exponential dispersion models read [2] and [3].

In spite of the distribution of the individual claim values being very well
represented in some situations by the gamma distribution, in other cases it could
be more suitable to attribute a lognormal distribution for Z1 , Z2 , ... . For instance,
in collision situations in car insurances and in common fires, where the individual
claim values can increase almost without limits but cannot fall below zero, with
most of the values near the lower limit and where the natural logarithm of the
individual claim variable yields a normal distribution.
(w)
NP
The aim of this paper is to estimate the parameters of Y (w) = Zi I(N (w)>0)
i=1
and N (w) distributions, where N (w), Z1 , Z2 , ... are independent, Z1 , Z2 , ... is a se-
quence of random variables with lognormal(µ, σ 2 ) distribution and N (w) follows
a Poisson distribution with rate λ.

Simulated examples are given to illustrate the methodology. The use of

a real dataset is not possible due to the high confidentiality with which the
companies deal with their database.
134 Zuanetti, Diniz and Leite

2. LOGNORMAL MODEL

A positive random variable Z is lognormally distributed if the logarithm of

the random variable is normally distributed. Hence Z follows a lognormal(µ, σ 2 )
distribution if its density function is given by
1  
2 (2π σ 2 )− 2 1
2
(2.1) fZ (z; µ, σ ) = exp − 2 log(z) − µ ,
z 2σ

for z > 0, −∞ < µ < ∞ and σ > 0.

The moments of the lognormal distribution can be calculated from the

moment generating function of the normal distribution and are defined as
 
k 1 2 2
(2.2) E[Z ] = exp kµ + k σ .
2

Thus, the mean of the lognormal distribution is given by

 
1 2
(2.3) E[Z] = exp µ + σ
2

and the variance is given by

(2.4) Var[Z] = exp 2µ + 2σ 2 − exp 2µ + σ 2 .

Products and quotients of lognormally distributed variables are themselves

lognormally distributed, as well as Z b and bZ, for b 6= 0 and Z following a
lognormal(µ, σ 2 ) distribution ([1]). However, the distribution of the sum of inde-
pendent lognormally distributed variables, that appears in many practical prob-
lems and describes the distribution of Y (w)|N (w), is not lognormally distributed
and does not present a recognizable probability density function ([7]).

Approximations for the distribution of the sum of lognormally distributed

random variables are suggested by Levy ([5]) and Milevsky and Posner ([6]).

3. PARAMETER ESTIMATION

As described in the previous section, the distribution function for Y (w),

where the claims Zi are independently and identically lognormal(µ,σ 2 ) distributed,
is not known. Consequently, the joint distribution for (N (w), Y (w)) and the cor-
responding likelihood function for the parameters µ, σ 2 and λ cannot be exactly
defined.
A Lognormal Model for Insurance Claims Data 135

However, since the lognormal distribution was defined with reference to the
normal distribution, estimate µ, σ 2 and λ from the likelihood function for these
parameters considering the variables N (w) and Y (w) is equivalent to estimate µ,
σ 2 and λ from the likelihood function based on the variables N (w) and
N (w)
X
X+ (w) = Xi I(N (w)>0) ,
i=1

where N (w) follows a Poisson(wλ), Xi = log(Zi ) follows a Normal(µ,σ 2 ) and the

Zi ’s are independent identically lognormal(µ,σ 2 ).

Then, we have
(3.1) X+ (w)|N (w) = n ∼ Normal(nµ, nσ 2 ), for n ≥ 1 .

The joint density of X+ (w) and N (w), for n ≥ 1, is defined as

f(X+ (w),N (w)) x+ , n; µ, σ 2 , λ =
(3.2)  
(wλ)n 1 2
= √ exp − (x+ − nµ) − wλ I(0,∞) (x+ )
n! 2πn σ 2 2n σ 2
and


(3.3) f(X+ (w),N (w)) x+ , 0; µ, σ 2 , λ = exp{−wλ} I(0,∞) (x+ ) .

In this work, without loss of generality, w is assumed to be equal to 1.

Considering (x+1 , n1 ), (x+2 , n2 ), ..., (x+m , nm ) observations from the independent
random vectors (X+1 , N1 ), (X+2 , N2 ) ..., (X+m , Nm ), where Ni ∼ Poisson(λ),
X+i | (Ni = ni ) ∼ Normal(ni µ, ni σ 2 ), i = 1, 2, ..., m, and m is the number of groups
present in the portfolio and considering δi = 0 for Ni = 0 and δi = 1 for Ni > 0,
the log likelihood function for the parameters µ, σ 2 and λ is given by
(3.4) l(µ, σ 2 , λ) =
m
(   )
X 1
1
2
= δi − log 2πni σ 2 + ni log(λ)− x+i − ni µ − λ + (1−δi ) (−λ)
2 2ni σ 2
i=1

If σ 2 = σ02 is known the maximum likelihood estimates of µ and λ are given

by
P
m P
m
δi X+i X+i m
X
∧ i=1 i=1
(3.5) µ = = if Nj > 0 ,
P
m P
m
δj Nj Nj j=1
j=1 j=1

and
P
m P
m
∧
δi Ni Ni
i=1 i=1
(3.6) λ = = .
m m
136 Zuanetti, Diniz and Leite

P
m Pm
Let S = Ni be the total number of claims and U = X+i . Hence,
i=1 i=1
 Pm P
m 
S follows a Poisson(mλ) and U | (N = n) follows a Normal µ ni , σ02 ni ,
i=1 i=1
where N = (N1 , N2 , ..., Nm ) and n = (n1 , n2 , ..., nm ) is the observed vector of
number of claims for m groups. Thus U | (S = s) follows a Normal(µs, σ02 s) and
∧
the exact distribution of λ is given by
 
∧ c
exp(−mλ) (mλ)c
(3.7) P λ= = P S =c = for c = 0, 1, 2, ... .
m c!

∧
The cumulative distribution function of µ given S > 0, F∧ (v), for v ∈ R
µ|S>0
is " #
∧  ∧
∞
[

P µ ≤ v|S > 0 = P µ≤v ∩ S = j |S > 0
j=1
!
∧ S
∞

P µ ≤ v, S =j , S > 0
j=1
=

P S>0
!
∧ S
∞

P µ ≤ v, S =j
j=1
=

P S>0
P
∞ ∧ 
P µ ≤ v, S = j
j=1
=

P S>0
P
∞ ∧ 

P µ ≤ v|S = j P S = j
j=1
(3.8) =

P S>0
P
∞  

P US ≤ v|S = j P S = j
j=1
=

P S>0
P
∞  

P Uj ≤ v|S = j P S = j
j=1
=

P S>0
P
∞  

P U ≤ jv|S = j P S = j
j=1
=

P S>0
X∞
exp(−mλ) (mλ)j
= FU (jv)
,
j=1
j! 1 − exp(−mλ)

where FU is the cumulative distribution function of the Normal(µj, σ02 j) distri-

bution, for j = 1, 2, ... .
A Lognormal Model for Insurance Claims Data 137

The corresponding probability density function is defined as

dF∧ (v)
µ|S>0
f∧ (v) =
µ|S>0 dv
∞
X exp(−mλ) (mλ)j
= fU (jv)
j
j=1
j! 1 − exp(−mλ)
∞
exp(−mλ) X (mλ)j
(3.9) = fU (jv)
1 − exp(−mλ) (j −1)!
j=1
∞
exp(−mλ) X
(mλ)r+1
= fU (r+1)v
1 − exp(−mλ) (r)!
r=0
∞
(mλ) exp(−mλ) X
(mλ)r
= fU (r+1)v ,
1 − exp(−mλ) (r)!
r=0

where fU is the probability density function of the Normal µ(r+1), σ02 (r+1)
distribution.

Let k be the number of groups with number of claims greater that zero.
If σ 2 is unknown, the maximum likelihood estimate of σ 2 is

 
P
m
(X+i − Ni µ )
∧ 2

∧ δi Ni
2 i=1
σ =
P
m
δi
i=1
(3.10)  
P
k (X+j − Nj µ )
∧ 2

Nj
j=1
= if Nj > 0, for all j = 1, 2, ..., k .
k

Using the invariant principle of maximum-likelihood estimation, the esti-

mates of E[Z], Var[Z], E[N ] and Var[N ], where Z represents the individual claims
and N the number of claims, are, respectively

 
b ∧ 1 ∧
E[Z] = exp µ + σ 2 ,
2
 ∧  ∧
∧ ∧
d
Var[Z] = exp 2 µ + 2 σ − exp 2 µ + σ 2 ,
2

∧
b ] = λ
E[N
and
∧
d ] = λ.
Var[N
138 Zuanetti, Diniz and Leite

4. THE LOCATION PARAMETER µ AS A FUNCTION OF

A COVARIATE

Suppose that (x+1 , n1 ), (x+2 , n2 ), ..., (x+m , nm ) are observations of the in-
dependent random vectors (X+1 , N1 ), (X+2 , N2 ) ..., (X+m , Nm ), m is the number
of groups present in the insurance portfolio, Ni ∼ Poisson(λ), and X+i | (Ni = ni )
∼ Normal(µi , ni σ 2 ), i = 1, 2, ..., m, with the following regression structure for the
location parameter
Xni
µi = α ni + β vij ,
j=1
where vij represents the covariate of the j-th individual claims of the i-th group,
for i = 1, 2, ..., m and j = 1, 2, ..., ni .
P
ni
Defining ri = vij , the log likelihood function for the parameters α, β, σ 2
j=1
and λ is given by


l α, β, σ 2 , λ =
m
(   )
X 1
1
2
2
= δi − log 2πni σ + ni log(λ) − x+i − α ni − β ri + (−λ) .
2 2ni σ 2
i=1

Let k, k ≤ m, be the number of groups with the number of claims greater

P
k
that zero, so that Nj > 0. The maximum likelihood estimates of α, β, σ 2 are
j=1
obtained through the data of only these k groups and are given by
P
k ∧ P
k
X+j − β rj
∧ j=1 j=1
(4.1) α = ,
P
k
Nj
j=1
PX P r
k k

P
k +j j

PN
X+j rj j=1 j=1
Nj − k
j=1 j
∧ j=1
(4.2) β = 
P r
k 2 ,
P
k rj2
j

PN
j=1
Nj − k
j=1 j
j=1
 
P
k
(X+j − µj )
∧ 2

∧ Nj
j=1
(4.3) σ2 = , Nj > 0, for all j ,
k
∧ ∧ ∧
where µj = α nj + β rj .

The maximum likelihood estimates of λ is defined as (3.6).

A Lognormal Model for Insurance Claims Data 139

5. APPLICATIONS

In order to illustrate the methods outlined in this article, two simulated

data set, with 20 insurance groups each, are presented. For the i-th group we
generated one observation of N following Poisson(λ) and ni observations of Z
following lognormal(µ, σ 2 ) and we considered Xj = log(Zj ) for j = 1, 2, ..., Ni and
PNi
X+i = Xj for each group. These observations, together with the values of N ,
j=1
are used in the estimation of the parameters. The first data set was simulated
considering a small rate of occurrence of claims in each insurance group and,
consequently, a large probability of groups with zero claims. The second data set
was simulated considering a large rate of occurrence of claims and, consequently,
a small number of groups with zero claims. In both cases the values of µ and σ 2
considered in the simulation of the data was 7.1 and 0.1, respectively. Thus

E[Z] = 1274.11 and Var[Z] = 170728.8 ,

that is, the expected individual claim value is 1274.11 MU with a variance of
170728.8 MU.

5.1. Portfolio with small rate of occurrence of claim

Considering m = 20, N ∼ Poisson(2), and the Zi ’s iid lognormal(7.1, 0.1),

we have  
P N = 0 = exp(−2) = 0.135 ,
that is, the probability of occurrence of no claims in each group is equal to 0.135.

The simulated individual claim values vary between 634.48 MU and

2819.6 MU and the observed values of N , X+ and δ are presented in Table 1.
Note that four of the twenty groups have no occurrence of claims.

Table 1: Observed values of N , X+ and δ for a simulated insurance portfolio

N X+ δ N X+ δ N X+ δ

1 1 6.79 1 8 2 14.63 1 15 3 21.58 1

2 3 21.12 1 9 1 6.89 1 16 0 0.00 0
3 0 0.00 0 10 1 7.29 1 17 2 13.54 1
4 3 20.98 1 11 1 6.54 1 18 2 14.48 1
5 2 13.56 1 12 3 21.97 1 19 0 0.00 0
6 0 0.00 0 13 1 7.03 1 20 2 14.24 1
7 3 21.82 1 14 4 27.69 1 Total 34 240.17 16
140 Zuanetti, Diniz and Leite

The estimates of λ, µ and σ 2 , calculated by (3.6), (3.5) and (3.10), respec-

tively, as well as a comparison between the true values of the parameters and
their estimates are presented in Table 2.

Table 2: The parameters true values and their estimates

True value Estimate Difference

λ 2 1.7 0.3
µ 7.1 7.06 0.04
σ2 0.1 0.09 0.01

∧
From the distribution function of µ given S > 0, defined in (3.8), we can
∧

calculate P µ ≤ v|S > 0 for different values of v ∈ R. Table 3 shows the prob-
∧
ability of µ ≤ v given S > 0, considering λ = 2, µ = 7.1, σ 2 = 0.1 (used for the
data simulation) and s = 34 (observed in this dataset).

∧

Table 3: P µ ≤ v|S > 0 for λ = 2, µ = 7.1, σ 2 = 0.1 and s = 34

∧
∧
∧

v P µ ≤ v|S > 0 v P µ ≤ v|S > 0 v P µ ≤ v|S > 0

4 0.0013 7 0.8063 10 0.9960

4.25 0.0060 7.25 0.8529 10.25 0.9974
4.5 0.0193 7.5 0.8913 10.5 0.9980
4.75 0.0484 7.75 0.9208 10.75 0.9986
5 0.0980 8 0.9424 11 0.9990
5.25 0.1721 8.25 0.9585 11.25 0.9993
5.5 0.2656 8.5 0.9709 11.5 0.9995
5.75 0.3687 8.75 0.9804 11.75 0.9996
6 0.4718 9 0.9863 12 0.9997
6.25 0.5787 9.25 0.9892 12.25 0.9998
6.5 0.6635 9.5 0.9925 2.5 0.9998
6.75 0.7473 9.75 0.9952 12.75 0.9999

Note that, from the results of Table 3,

 ∧ 
P 4.5 ≤ µ ≤ 9.25 = 0.9699 .

5.2. Portfolio with large rate of occurrence of claim

In the second dataset, twenty observations of N were generated from the

Poisson(100) distribution and the Zi ’s were generated from the lognormal(7.1, 0.1)
distribution. Consequently,
 

P N = 0 = exp −100 ≃ 0 ,
A Lognormal Model for Insurance Claims Data 141

that is, the probability of occurrence of no claims in each group is practically null.
The observed values of N , X+ and δ are presented in Table 4 and the simulated
individual claim values vary between 440.91 MU and 3212.9 MU.

Table 4: Observed values of N , X+ and δ for a simulated insurance portfolio

N X+ δ N X+ δ N X+ δ

1 95 676.0 1 8 86 606.6 1 15 79 553.4 1

2 104 739.3 1 9 108 762.0 1 16 95 674.5 1
3 85 601.6 1 10 85 601.1 1 17 87 619.8 1
4 92 652.6 1 11 98 695.8 1 18 105 747.6 1
5 106 749.6 1 12 86 612.0 1 19 101 717.3 1
6 111 791.9 1 13 83 586.1 1 20 100 714.6 1
7 85 600.8 1 14 100 709.8 1 Total 1891 13412.5 20

The estimates of λ, µ and σ 2 , calculated by (3.6), (3.5) and (3.10), respec-

tively, as well as a comparison between the true values of the parameters and its
estimates are displayed in in Table 5.

Table 5: The parameters true values and their estimates

True value Estimate Difference

λ 100 94.55 5.45

µ 7.1 7.093 0.007
σ2 0.1 0.096 0.004

∧
Table 6 shows the probability of µ ≤ v given S > 0, considering λ = 100,
µ = 7.1, σ 2 = 0.1 (used for the data simulation) and s = 1891 (observed in this
dataset).

∧

Table 6: P µ ≤ v|S > 0 for λ = 100, µ = 7.1, σ 2 = 0.1 and s = 1891

∧
∧
∧

v P µ ≤ v|S > 0 v P µ ≤ v|S > 0 v P µ ≤ v|S > 0

6.1 0.0000 6.6 0.2216 7.1 0.9929

6.2 0.0001 6.7 0.4639 7.2 0.9989
6.3 0.0019 6.8 0.7152 7.3 0.9999
6.4 0.0150 6.9 0.8873 7.4 1.0000
6.5 0.0722 7.0 0.9671 7.5 1.0000

∧
From the results of Table 6 we have P [6.4 ≤ µ ≤ 7.1] = 0.978.
 142 Zuanetti, Diniz and Leite

6. CONCLUDING REMARKS

The theory for exponential dispersion models cannot be applied to estimate

the parameters µ, σ 2 , that specify the lognormal distribution of the individual
claim value (Z), and λ, the occurrence rate of claims, because the lognormal dis-
PN (w)
tribution and, consequently, the joint distribution of Y (w) = i=1 Zi I(N (w)>0)
and N (w) does not belong to the class of the exponential dispersion model.
NP(w)
However, from the joint distribution of X+ (w) = log(Zi ) I(N (w)>0) and N (w),
i=1
maximum likelihood estimates of µ, σ 2 and λ can be defined and applied to an
insurance portfolio dataset, in which N (w) follows a Poisson(wλ) distribution
and Z is lognormally distributed.

ACKNOWLEDGMENTS

We would like to thank the associate editor and referees for their valuable
suggestions and comments on an earlier version of this paper.

REFERENCES

[1] Crow, E.L. and Shimizu, K. (Eds.) (1988). Lognormal Distributions: Theory
and Application, Dekker, New York.
[2] Jorgensen, Bent (1986). Some properties of exponential dispersion models,
Scand. J. Statist., 13, 187–198.
[3] Jorgensen, Bent (1987). Exponential dispersion models (with discussion),
J. Royal Statistical Society Ser. B, 49, 127–162.
[4] Jorgensen, B. and Souza, M.C.P. de (1994). Fitting Tweedie’s compound
poisson model to insurance claims data, Scandinavian Actuarial Journal, 69–93.
[5] Levy, E. (1992). Pricing European average rate currency options, Journal of
International Money and Finance, 14, 474–491.
[6] Milevsky, M.A. and Posner, S.E. (1998). Asian Options, the sum of lognormal,
and the Reciprocal Gamma distribution, Journal of Financial and Quantitative
Analysis, 33, 3.
[7] Slimanc, S.B. (2001). Bounds to the distribution of a sum of independent log-
normal random variables, IEEE Transations on Communications, 20, 6.
[8] Tweedie, M.C.K. (1947). Functions of a statistical variate with given means,
with special reference to Laplacian distributions, Proc. Camb. Phil. Soc., 49,
41–49.

View publication stats