Re: SPRING 98 Exam

David Kennerud ( kennerud@concentric.net )
Tue, 20 Oct 1998 19:46:21 -0700

Carrie,

Hope this helps. Let me know if you have trouble understanding my notation
or solutions.

If anyone has simpler solutions to any of these, I'd love to see them.
(Especially #13 and #22).

Dave Kennerud
kennerud@concentric.net

#2) Given lambda (I'll use l for lambda), HM = l. So VHM = Var(l) = u
Similarly, given l, PV = l. So EVPV = E(l) = u.
So k = EVPV / VHM = u / u = 1.

Answer: A

#4) I'll use l for lamda, and I'll use a and b for the gamma parameters (a
= alpha and b = lambda as in Hogg and Klugman)

First, calculate a and b: a / b = mean of gamma = 0.05
a / (b^2) = var. of gamma = 0.01
Solving for a and b yields: a = 0.25 b = 5

The posterior distribution will also be gamma (always the case for
Poisson/gamma problems), with parameters
a' = a + number of claims = 0.25 + n
b' = b + number of exposures = 5 + 10 = 15

Since the posterior variance = prior variance: a' / (b'^2) = 0.01
0.25 + n = 0.01 * 15^2 = 2.25
n = 2

Answer: C

#13) The probability of a single claim being less than k is: F(k) = 1 - e^(-k)
So if there are n claims, the prob. all are less than k is: (1 - e^(-k))^n
The probability of n claims is p(n) as given: p(n) = (e^l - 1)^(-1) *
(l^n) / n!
So the prob. of having exactly n claims, all < k is: (e^l - 1)^(-1) *
(l^n) / n! * (1 - e^(-k))^n
= (e^l - 1)^(-1) * { [ l * (1 - e^(-k) ]^n / n! }
(This is equivalent to the probability of exactly n claims, the largest
of which is less than k)

Now, to get the prob. that the largest claim is less than k, you need to
sum the above over all n = 1 to infinity.
Forget about the (e^l - 1)^(-1) for the moment, since that has no n's in it.

Let y = [ l * (1 - e^(-k) ]. So [ l * (1 - e^(-k) ]^n / n! = y^n / n!,
and what you need to sum is:

y^n / n! n = 1 .. infinity
But summing y^n / n! from 0 to inf. is e^y, so summing from 1 to inf. is
(e^y - 1) = (e^[ l * (1 - e^(-k) ] -1)

So the final answer (remember the e^l - 1 ) is: (e^l - 1)^(-1) * (e^[ l *
(1 - e^(-k) ] -1)

Answer: D

#20) Calculate the MLE for l.
L(x1,...,xn) = l^(-n) * e^(- (x1 + . . . + xn) / l )
ln(L) = -n * ln(l) - (x1 + . . . + xn) / l )
d(ln(L))/dl = -n/l + (x1 + . . . + xn)/l^2
Setting this to 0 and solving for l yields: l = Sample mean (I'll use X
for sample mean).
So e^(-1/l) is apx. e^(-1/X).

Answer: B

#21) Since there is only one comparison point, the minimum distance
statistic is just:

( F(1) - Fn(1) )^2

Since this has only one term, to minimize it all you need to do is to set
F(1) = Fn(1)

But F(1) = 1 - e^(-1/l), and Fn(1) = 1 - p, so: F(1) - Fn(1) = (e^(-1/l)) - p

So e^-(1/l) = p.

Answer: C

#22) We need to estimate the variance of: h(l) = e^(-1/l)
To do this you need the formula for the approximate variance of a function
of a parameter (p. 93 of H&K)
Since there is only one parameter, the apx. variance will be: h'(l)^2 * s^2,
where s is the apx. variance of the parameter l.

To calculate s^2, need to use Rao-Cramer: ln(f(x)) = -ln(l) - x/l

Taking 2 derivative with respect to l yields: 1/l^2 - 2*x/l^3

Taking the expected value (w.r.t. x): 1/l^2 - 2*E(x)/l^3 = 1/l^2 -
2*l/l^3 = -1/l^2

So s^2 will be: -1/(n* -1/l^2) = l^2 / n

Now, h'(l) = 1/l^2 * e^(-1/l)

So, finally, h'(l)^2 * s^2
= [1 /l^2 * e^(-1/l) ] ^ 2 * ( l^2 / n )
= 1/(n * l^2) * e^(-2/l)

Answer: D

#23) If the sample r.v. are xi, let yi = { 1 , if xi > 1
{ 0 , if xi <= 1

Then the yi are Bernouilli r.v. with prob. of success = prob(xi > 1) =
e^(-1/l)
Also, p = (proportion of xi greater than 1) = (y1 + . . . + yn)/n, so p
is the average of n Bernouilli r.v.

Since var(yi) = e^(-1/l) * (1 - e^(-1/l)), the variance of p will be:
e^(-1/l) * (1 - e^(-1/l) ) /n

Answer: E

#26) First, calculate the following assuming m is fixed. (N = claim ct
r.v., S = claim size r.v.)
E(N) = m; Var(N) = m; E(S) = 20m; Var(S) = 400m^2
Then the PV|m = E(N)*Var(S) + Var(N)*E(S)^2 = m*400m^2 + m*(20m)^2 = 800m^3

So EVPV = 800*E(m^3)

Now, E(m^3) = integral from 0 to infinity of m^5*e^(-m)/2. Since
m^5*e(-m)/(5!) is the pdf of a gamma r.v. with parameter lambda = 1 and
alpha = 6, this integral will be equal to: 5! / 2 = 60.

Thus, EVPV = 800*60 = 48,000

Answer is D

#29) The posterior p.d.f. will be proportional to the Likelihood function,
which is:

g(l) * f(50) = 500,000*(1/l^5)*e^(-150/l)

The only solution that has this form ( C * (1/l^5) * e(-150/l) for some
constant C) (Note that you don't actually have to find the constant C)

Answer. E

>X-Lotus-FromDomain: SUNLIFE
>From: Carrie_Cheung/Actuarial/Corporate/SunLife@SunLifeOfCanada.com
>To: studygroup4b@lists.casact.org
>Date: Mon, 19 Oct 1998 09:26:39 -0400
>Subject: SPRING 98 Exam
>
>Hi,
>
>Everyone must be study hard. I wonder whether there is any one who want to
>try the SPRING 98 Exam. I have problem with quiet a number of questions.
>
>Question 2, 4, 13, 20-23( they are multiple question, which I have no where
>to start ), 26 and 29.
>
>If some one prefer to fax their work ( instead of typing) , my fax number
>is (416) 979-6111. Or if you have problems with questions other than the
>above, I might be able to help you.
>
>Thanks and Good Luck to your Exams,
>Carrie
>
>
>
>