Actuarial Society of India

EXAMINATIONS

May 2003


Subject 106 Actuarial Mathematics II


Indicative Solution



1. Let us find the probabilities of guessing right or wrong.
If 5, the probability that the guess is right,
2 3 3
5
5
3 3
3 125
P[ ] [ ]
125
k
k
x x k
X k dx

>

.

If 3, the probability that the guess is right is,
2 3 3
0 3 3
0
3
P[ ] [ ]
27
k
k
x x k
x k dx

<

.

The risk function will be calculated from the size of loss multiplied by the probability that
the guess is wrong. So, for this decision rule adopted by the statistician
3
[ , 5] 40
125
k
R d , and
3 3
[27 ]
[ , 3] 90 [1 ] 90
27 27
k k
R d

.

The maximum risk will be minimized when the two expressions will have the same value.
(i.e)
3 3
3 3
3
3
[27-k ]
40 90
125 27
, 12 125[27 ]
or, 137 125 27 3375
or, 24.635,
or, 2.9097
k
or k k
k
k
k

[3]








[3]












[2]

2.












The effect of 10% inflation will be to change the mean and standard deviation of the
claim size distribution to Rs. 11,000 and Rs. 6,600 respectively, in the following year.
The companys retention limit will become Rs.55, 000.
Because of the multiplicative property of the lognormal distribution, next year claim size
will be lognormal.
Given,

2
exp[ ] 11000;and
2

+ ---------------------------------------(1)
( )
2 2 2
Varience exp 2 [exp( ) 1] (6600) + g ----------(2)
Dividing the second equation by the square of the first equation, we get


























(i)















(ii)











(iii)








[ ]
2
2
2
2
6600
exp 1 0.6 0.36
11000

]
]
] ]
.

Or,
2
ln1.36 0.307485 .
Or, 0.554513 .
Substituting the value of
2
in the first equation, we have

[ ] exp 0.1537425 11000
or 0.1537425 ln11000 9.30565
9.30565 0.15374 9.15191. or

+
+

.

The entire loss will be borne by the insured if the full claim amount is less than Rs.
5,000.
[ ]
( )
ln ln5000 9.15191
P 5000 , i.e.P
0.554513
ln 8.51719 9.15191
i.e. P
0.554513
ln
i.e. P 1.145
1 1.145 1 0.8739 0.1261 12.6%
X
X
X
X

]
< <
]
]
]
<
]
]
]
<
]
]


i.e. one claim in eight.

The reinsurer will make the payment only when the claim amount exceeds Rs. 55000.
So, the proportion of such claims will be,
[ ]
( )
ln55000 9.15191
P X>55000 1
0.554513
1 3.180
1 0.99926
0.00074.

| `


. ,

So, the proportion is about 0.07%

The amount of X
ph
paid by the insured himself is

5000
5000 5000
{
ph
X if X
X
if X
<

The average amount paid by the insured himself is
5000
0 5000
E ( ) 5000 ( )
ph
X xf x dx f x dx

] +
]












[2]














[2]











[2]






































(iv)




















where ( ) f x is the p.d.f. of the lognormal distribution.
( )
2 1
2
2
ln5000
First integral
8.51719 9.15191 0.307485
11000
0.554513
11000 1.699
11000[1 0.95534] 11000 0.04466
491.27
491.30,say.
e

+
| `


. ,

| `


. ,

The second integral
5000 P[ 5000]
5000[1 0.1261] from(1)above.
=5000 0.8739
=4369.5
X >


Adding we get Rs. 491.3 + Rs. 4,369.5 = Rs. 4,860.80

The amount paid by the direct insurer will be

X
direct
={
0 if 5000
5000 if 5000< 55000
50000 if 55000
X
X X
X

<

The average amount paid by the direct insurer will be:

[ ] ( ) ( ) ( )
( ) ( ) ( )
( )
55000
5000 55000
55000 5000 55000
0 0 0
5000
0
55000
E 5000 50000
5000
5000 ( ) 50000 ------(1)
direct
X x f x dx f x dx
xf x dx xf x dx f x dx
f x dx f x dx

+
+
+




But
















[1]










[2]








































































[ ]
55000
0
ln55000 9.15191 0.307485
xf(x)dx = 11000
0.55453
10.915088 9.15191 0.307485
11000
0.55453
11000 2.6251
11000 0.99567
10,952.37
]

]
]
]

]
]

[ ]
5000
0
ln5000 9.15191 0.307485
xf(x)dx = 11000
0.55453
11000 1.6991
11000 (1 0.99567)
491.27

]

]
]

[ ]
55000
0
ln55000 9.15191
5000 f(x)dx = 5000
0.55453
5000 3.1796
5000 0.99926
4996.3

]
]

[ ]
5000
0
ln5000 9.15191
5000 f(x)dx = 5000
0.55453
5000 1.1446
5000 0.1261
630.5

]
]

[ ]
[ ]
55000
50000 f(x)dx = 50000 1 (3.1796)
50000 1 0.99926
50000 0.00074
37

(v)




(vi)

[ ]
Substituting these in(1)
E 10,952.37 491.3 4996.3 630.5 37 6132.3
direct
X + +

= Rs. 6132.3


The average amount paid by the reinsurer in respect of all claims

.11000 .4860.8 .6132.3
.6.90
Rs Rs Rs
Rs

The average amount paid by the reinsurer in respect of claims in which he is involved is
the conditional mean
[ ]
E | 0 i.e. the average amount for non-zero claims
re re
X X > .
This can be calculated from the results obtained before:

[ ]
[ ]
[ ]
E
6.90
E | 0
P 55000 0.00074
.9324.32
re
re re
X
X X
X
Rs
>
>

[4]




[1]










[2]

3.
(a)


























The assumptions underlying the Basic chain-ladder method are:
(i) The chain-ladder method technique is based on the assumption that payments
from each accident year will develop in the same way. In other words, the
same development factors can be used to project outstanding claim for each
accident year.


(ii) It is assumed that the weighted average of the past inflation will be repeated in
the future.

Let us calculate the development factors:

0,1
1,2
2,3
542 565 644
1.13114
480 500 568
576 602 1178
1.06414
542 565 1107
590
1.02431
576
f
f
f
+ +

+ +
+

+


We shall use the above development factors to project future claims.














[2]










[2]




















(b)


Development Year
0 1 2 3
Claims incurred Accident Year
1998 480 542 576 590
1999 500 565 602 616.63
2000 568 644 685.31 701.97
2001 550 622.13 662.03 678.12


Assuming that the claims are fully run-off by the development year 3, the total claims
payable is: 590 + 616.63 + 701.97 + 678.12 = 2586.72 (thousands).
So, the outstanding claims reserve is estimated as Rs. 2586.72 Rs. 2390
= Rs. 196.72 (thousands).
Assuming that in respect of year 1998 is fully run off, the estimated loss ratio is

590
0.85
694
.
Let us consider each of the accident years in turn.
Accident Year 1999
Initial estimate of the ultimate loss is 710 X 0.85 = 603.50
Expected claims incurred by the end of Development Year 2 is
2,3
603.50 603.50
589.20
1.02431 f

Actual claims incurred by the end of the Development Year 2 is 602.
The difference is 602 589.20 = 12.80.
So, we should increase our final estimate for the year to
Rs. 603.50 + 12.80 = 616.30.


Accident Year 2000
Initial estimate of ultimate loss = 850 X 0.85 = 722.50
Expected Claims
1,2 2,3
722.50 722.50
662.8
1.02431 1.06414 f f



Actual claims incurred up to the end of Development Year 1 is = 644.
The difference is 662.80 644.00 = 18.80.
So, we should reduce our estimate for the year to 722.50 18.80 = 703.70

Accident Year 2001
Initial estimate of the ultimate loss is 800 X 0.85 = 680.
Expected Claims incurred by the end of the Development Year 0 is
0,1 1,2 2,3
680 680 680
551.5
1.13114 1.06414 1.02431 1.23295 f f f



Actual Claims incurred by the end of the Development Year 0 is 550.
Difference between Actual claims and Expected claims = 1.5
Revised estimate for the ultimate loss is 680 1.5 = 678.5






[2]





[1]















[2]









[2]










Hence estimated total amount of ultimate loss is: 590 + 616.3 + 703.7 + 678.5 = 2588.5
The outstanding claim reserve is: 2588.5 2390.0 = 198.5 lakhs

The main assumption is that the estimated loss ratio is appropriate to estimate the
ultimate loss, in additions to the assumptions in the chain-ladder method.


[2]



[1]

4.
(i)







(ii)
(a)












(b)



















The moment generating function of S is
( )
( ) ( )
( ) {
( ) ( )
E E E |
E E
St St
S
N
Xt
N X
M t e e N
e G M t
]

]
]

]
]


Let N be the number of claims.
( ) ( )
( )
3
0
1 4
3
1
0 0
1
P P | 3
3 4
3
! ! 4
3
, 0,1,2,... integrand is Gamma 4, 1
4
n n n
n
n
N n N n e d
e e
e d d
n n
n n

+
+
]
]

+

g
Q



1
0 0
3 3 3
has pgf
4 4 4 4
n
n
n
s
N s
s

+
| `

. ,

.


For gamma (,), mean
2
is 2, Variance is 2, giving 2; 1.




So,
( )
( )
( )
( )
( )
( )
2
2
2
2
2
1
1 , 1
1
1
3 1
3
1
4 8 3
4
1
X
S
t
M t t
t
t
M t
t t
t

| `
<

. ,

[2]





[1]






[2]





[1]








[1]



[1]




5.















(i)
































In general, if the probability distribution is
+1
+1
Pareto( , ) and relation is M, the probability distribution
involving the reinsurer is
P(x>M) x>M , 0
( +x)
that is
1 1 for x>M ,
( +x) M

]
>
]
]
]
| `
]
| `
]
]

+
. ,
] ]
. , ]
1
>1
( +M)
= for x>0 , >1
( +M+x)
That is, this distribution is Pareto( +M, )

+


Using this, the likelihood function is given by,

( )
( ) ( ) ( )
5
1
i=1
5
1
(300000)
(300000 )
ln 5ln 5 ln300000 1 ln 300000
c
i
i
L
x
L x

+
+ + +

( )
( )
( )
5
1
5
1
ln 0, becomes
5
ln 300000 5ln300000
5
=
ln 300000 5ln300000
i
i
i
i
d
L
d
x
or
x

+
+

[2]















































(ii)






6.
(i)














(ii)
(a)





( )
5
1
ln 300000 ln385000 ln458000 ln318000
+ ln445000+ln390000
= 64.44516
and 5ln300000 = 63.05769
So that
5 5
=
64.44516-63.05769 1.38
i
i
x

+ + +

2
2 2
3.60368
747
5
0 so that given above is the
maximum likehood estimate of
d l
Also
d

<



The mean amount paid by the reinsurer is the mean of

300000

Pareto(250000+50000, )= 115221
3.60368-1

(Using the result on page 14 of the Actuarial Table )


( )
( )
( )
( )
5
1
1
5 5
5
1
|

=5
|
That is, | Gamma 1, 5
i
i
i
i
x
x n n
i i
i
i
x
n
x n
n
i
e e
f x
x
x
e
f e e
f x e
x x n

| `
+ +

. ,

g
:



Under quadratic loss, the Bayesian estimate of is the posterior mean.
( )
1
1
E |
5
n
i
x
x
n

+

+

[2]










[3]




[1]










[2]








[1.5]

(b)














(iii)





Under all-or-nothing loss, the Bayesian estimate is the posterior mode.

If x
i
= 0, then the mode is at =0.

Otherwise, the posterior density has a maximum when
( )
|
0
df x
d

That is,

( )
( )
( )
( )
1 5 5
5 0
or, =
5
i i
x x n n
i
i
x e n e
x
n

+ +

+
+

Under absolute error loss, the Bayesian estimate is the posterior median.
( )
( )
( )
( ) ( )
2 10
1
2
10 10
1
2
10
1
2
10
5, 1, so that | 2,10
10
Median is given by
2
, 10 10
. . 10 1
. . 2 10 1 0 , say.
i
m
m m
m
m
n x x gamma
e
d
or e e d
i e m e
i e m e g m

] +
]
+
+

:


m 2(1+10m) e
10m
g(m) Root
0.1 4 2.7 1.3 >0.1
0.2 6 7.4 - 1.4 <0.2
0.15 5 4.5 0.5 >0.15
0.16 5.2 4.95 0.25 >0.16
0.17 5.4 5.47 -0.07 <0.17
0.165 5.3 5.21 0.09 >0.165


Root of the equation lies between 0.165 and 0.17.
Hence m = 0.17, correct to two significant figures













[1.5]














[2]












[2]



7.
(i)




( ) ( ) ( )
4
3
P 0.8 0.2 , k=4; q=0.2; p=0.8
n
n
N n
n
+
| `


. ,
Q















(ii)
(a)



































(b)
( )
( )
( ) ( )
( ) ( )
( )
4
4 1
3
0.8 0.2
P 3 0.6
0.2 0.2
2 P 1
0.8 0.2
1
n
n
n
N n n n
n N n n n
n

+
| `

+
. ,
+
+ | `

. ,

The relationship holds with a = 0.2 and b = 0.6


Using the recursive method,
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( )
1
4
P P P , 1
P 0 P 0 0.8 0.4096.
(Working in units of 500) we have,
3 1
P 1 0.2 1 P 1 P 0
1
0.2 4 0.5 0.4096 0.16384
3 1 3 2
P 2 0.2 1 P 1 P 1 0.2 1
2 2
s
x
bx
S s a X x S s x s
s
S N
S X S
S X S

| `
+

. ,

| `
+

. ,

| ` | `
+ + +

. , .

( ) ( )
P 2 P 0
0.2 2.5 0.5 0.16384 0.2 4 0.25 0.4096 0.12288
X S

,
+
( ) ( ) ( ) ( ) ( )
3 1 3 2
P 3 0.2 1 P 1 P 2 0.2 1 P 2 P 1
3 3
0.2 2 0.5 0.12288 0.2 3 0.25 0.16384 0.049152
S X S X S
| ` | `
+ + +

. , . ,
+

( ) ( ) ( ) ( ) ( )
( ) ( )
3 1 3 2
P 4 0.2 1 P 1 P 3 0.2 1 P 2 P 2
4 4
3 4
0.2 1 P 4 P 0
4
0.2 1.75 0.5 0.049152 0.2 2.5 0.25 0.12288 0.2 4 0.25 0.4096
0.1058816
S X S X S
X S
| ` | `
+ + +

. , . ,

| `
+ +

. ,
+ +

Hence the probability that the aggregate claim amount is less than or equal to 2000 is
( )
P 4 0.4096 0.16384 0.12288 0.049152 0.1058816 0.8513536 S + + + +


Using the normal approximation,
( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
2
E E E and Var E Var E X Var S N X S N X N + ]
]







[2]







[1]











[2]














[2]

[1]




[1]
( ) ( )
2
4 0.2 4 0.2 5
E 1; Var
0.8 0.64 4
kq kq
N N
p p


Working in units of 1000, we have
( )
( )
( )
( ) ( )
( ) ( )
( )
2 2 2 2
2
2
E 0.5 0.5 1 0.25 2 0.25 1
E X 0.5 0.5 1 0.25 2 0.25 1.375
Var 1.375 1 0.375
5
E 1 1 1 and Var 1 0.375 1 1.625
4
2 1
P 2 P 0,1
1.625
= 0.7844645
= 0.783623
X
X
S S
S N
+ +
+ +

+


' '

[2]









[1]
8.
(i)





















(ii)
( ) ( ) ( ) ( )
( ) ( ) ( )
( )
( )
( ) ( )
( )
( )
( )
2
2
2
2
3
2
2 2
2
2 2
2
2 2
2 2
2 2
M E 1 E E ...
2!
So, M 1 E E , assuming that is small enough
2!
to omit etc. terms
= 1 E
2
1
2
. . ( )
2
2
2
,
tX
X
X
t
t e t X X
R
R R X X R
R
R
R X
R
cR RE X
R
i e c R
c
c
R R

+ + +
+ +
+ + +
]
+ + + +
]
]
+
| `

. ,

+
+
Q 0


The adjustment coefficient can be used to assess the effectiveness of different
reinsurance arrangements, using Lindbergs inequality
( )
, (where U is the initial surplus)
RU
U e

to find an upper bound for the
probability of ruin for the insurer under different reinsurance arrangements. An








[2]











[2]








arrangement that produces a lower upper bound for the probability of ruin is in some
sense more secure for the insurer than an arrangement that has a higher upper bound
for the probability of ruin. However the adjustment coefficient cannot tell anything about
the relative profitability of different reinsurance arrangements.



[3]


9. Let C be the aggregate claims paid in five years.
The distribution of claims is lognormal (,
2
), say.

As the number of policies sold increases each year by 10% p.a., the total number of policies sold
in five years is

5
100x[(1.1) 1]
610.51
0.1

The premiums collected will be 5000x610.51=3052550.
The expenses incurred will be 5000x0.3x610.51 915765.
Initial Surplus is Rs. 1000000.
Thus the surplus at the end of the fifth year will be

( )
5 1000000 3052550 915765 C U +
3136785 C
C is distributed as
2
ln( , ).
Mean of this distribution is
2
1
2
e
+
and standard deviation is
2
2
1
2
1 e e

+

Equating these with the given values,

2
1
2
e
+
= 0.6 x 5000 x 610.51 = 1831530, and

2
2
1
2
1 e e

+
= 1.8 x 5000 x 610.51 = 5494590.
Solving,
2
1 3, or 1.517427 and 13.26937 e

If, in the fifth year the insurer becomes insolvent, then

U(5)<0; The prob. for this is
Pr[C>3136785]

As C has lognormal distribution with parameters and ,
The required probability is
( )
ln3136785 13.26937
1
1.517427
1 1.1132924
1 0.8672
0.1328
13.28%
| `

. ,

[5]
10.
















(i) (a) A Covariate is a variable in a linear model about which we have some
information. We try to predict the values of the response variable by looking at the
values taken by the covariates. Covariates may take numerical values or categorical
values.
(b) A Linear Predictor is a function of the covariates that is used in the model. It is
the linear function of the parameters, whose value we try to estimate. It may not
necessarily be a linear function of the covariates themselves.

(c) A Link Function is the function that relates the response variable to a linear
function involving the covariates, in which case the link function will be the
identity function. In more usual examples we try to express some function of the
response variable in linear form. The function applied to the response variable is
called the link function. In many cases there is a natural form for the link function
to take. This is called the canonical link function.

(ii) Let the linear predictor be of the form
.
i j k
+ +
Step 1: Estimate the values of .
Using the above equation we get


Young Middle aged Old
Male Metro 0.838 0.227 0.359
Town -0.831 -1.442 -1.310
Countryside -1.326 -1.937 -1.805
Female Metro 0.598 -0.013 0.119
Town -1.071 -1.682 -1.550
Countryside -1.566 -2.177 -2.045


Step 2: Estimate value of


We have assumed that the underlying probability distribution is binomial. The canonical
link function associated with binomial distribution is :

ln , where is theclaimfrequency.
1

| `

. ,

Thus the estimated value of is given by

( )
1
1 e

+
Using this we get the estimated value of for each of the cells as shown below:


Young Middle aged Old
Male Metro 0.6980 0.5565 0.5888
Town 0.3034 0.1912 0.2125
Countryside 0.2098 0.1260 0.1412




[1]


[1]






[2]


































Female Metro 0.6452 0.4968 0.5297
Town 0.2552 0.1568 0.1751
Countryside 0.1728 0.1018 0.1146


(a) From the above Table we find that the maximum claim frequency is attributable
to young, male drivers operating in metropolitan area and it is estimated to be
0.698.
Similarly, minimum claim frequency is attributable to middle aged, female
drivers operating in countryside and it is estimated to be 0.1018.


(b) The residuals are the difference between the observed values and the estimated
alues. Thus we get,


Young Middle aged Old
Male Metro -0.017 0.0065 0.0022
Town 0.0156 -0.0042 -0.0045
Countryside 0.0002 0.0090 0.0088
Female Metro 0.0568 0.0152 -0.0237
Town -0.0102 -0.0108 -0.0001
Countryside -0.0018 0.0072 -0.0166


(c) For drivers having the same characteristics and operating in similar areas their
claim frequency may well be modeled using a binomial distribution. Thus the
assumption of binomial distribution is appropriate.

Any link function used must be differentiable and invertible. Obviously,

ln
1

| `

. ,
is differentiable and is also invertible as
( )
1
1 e

+ .


The values taken by the parameters are both positive and negative. Thus, apparently, the
linear predictor could take any value between and + .
But the claim frequency can take only values between 0 and +1. This is done by the link
function. Thus the link function gives sensible values and hence is appropriate.







[5]



[1]


[1]











[1]

















[3]




THE END