Faculty of Actuaries Institute of Actuaries

EXAMINATION
September 2007

Subject CT6 — Statistical Methods

Core Technical

EXAMINERS’ REPORT

Introduction

The attached subject report has been written by the Principal Examiner with the aim of
helping candidates. The questions and comments are based around Core Reading as the
interpretation of the syllabus to which the examiners are working. They have however given
credit for any alternative approach or interpretation which they consider to be reasonable.

M A Stocker
Chairman of the Board of Examiners

December 2007

© Faculty of Actuaries
© Institute of Actuaries
 Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

Comments

Overall it is clear that candidates found this a tougher paper than in other recent sittings. In
particular, candidates who did not have a firm grasp of the basic statistical material covered
in subject CT3 struggled with some of the questions.

Comments on individual questions are as follows:

Q1 The Examiners had intended this to be a straightforward bookwork question as it is

taken largely verbatim from the core reading. However, very few candidates could
recall this part of the core reading accurately and therefore most candidates
struggled to score any marks.

Q2 Well answered. There was a minor typographical error in the question, but virtually
all candidates seemed to have understood what was intended. The Examiners
allowed any alternative interpretation provided it was clearly stated.

Q3 In general, this question was very poorly answered. Most candidates appeared
unable to apply Bayes’ Theorem to the situation given in the question.

Q4 Well answered.

Q5 Parts (i) and (iii) were generally well answered, a pleasing improvement relative to
similar questions from recent sittings. Only the better candidates were able to
complete part (ii).

Q6 This was the first time in a number of sittings that this material has been tested. Most
of the better candidates coped reasonably well with parts (i), (ii) and (v). Weaker
candidates struggled badly with (i) and (ii) which required only some calculus. Part
(iii) was actually simpler than many candidates appear to have expected and as a
result many solutions were over-complicated and scored poorly.

Q7 Well answered.

Q8 This question was well answered by the better candidates. Many candidates picked
up significant follow through marks in parts (ii) and (iii) despite making numerical
errors in part (i).

Q9 Part (i) was generally well answered. The remaining parts were tougher, with a
number of candidates quoting the results in parts (iii) and (iv) rather than deriving
them as instructed.

Q10 Parts (i) and (iii) were well answered. Most candidates were able to make some
progress with (ii) (a) though a number struggled to derive three correct equations.
Only the best candidates were able to answer (ii) (b).

Page 2
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

1 An ARCH(p) model is

p
Xt = μ + et α0 + ∑ α k ( X t −k − μ) 2
k =1

where et are independent N(0,1).

Often, this is used to model ln(Zt/Zt-1) where Zt is the asset price.

It can be seen that a large departure in Xt-k from μ will result in Xt having a larger
variance. This will then result in a large volatility for the asset price.

2 (i) θ1 would dominate θ2 (and vice versa) if the amount of birds poached under
each of a1, a2 and a3 is higher in each case for θ1 (θ2)

θ1 provides a better outcome for a2 (90 > 0) and a3 (120 > 75)

θ2 provides a better outcome for a1 (0 < 75)

(ii) Minimax solution for gamekeepers. Worst loss under each option would be:

a1 = max (0,75,120) = 120

a2 = max (90,0,90) = 90
a3 = max (120,75,0) = 120

Therefore, gamekeepers would choose a2 and lose 90 birds.

(iii) a1 = 0.25 × 0 + 0.35 × 75 + 0.4 × 120 = 74.25

a2 = 0.25 × 90 + 0.35 × 0 + 0.4 × 90 = 58.5
a3 = 0.25 × 120 + 0.35 × 75 + 0.4 × 0 = 56.25

Hence the Bayes decision is a3.

Page 3
 Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

P ( N − r = k r big claims) = P ( N = r + k ) × P (of r + k claims k are small) / P (r big claims)

but

∞
λr+ j (r + j )! r
P (r big claims) = ∑ e −λ × p (1 − p ) j
j =0 (r + j )! r ! j !

λ r r ∞ −λ λ j
= p ×∑e (1 − p) j
r! j =0 j!

λ r r −λ λ (1− p )
= p e e
r!

So

P ( N − r = k r big claims) = P ( N = r + k ) × P (of r + k claims k are small) / P (r big claims)

λ r +k (r + k )! r
e −λ × p (1 − p ) k
(r + k )! r !k !
=
e −λ eλ (1− p ) p r λ r / r !

λ k (1 − p )k
= e −λ (1− p )
k!

which is a probability from a Poisson distribution with parameter λ (1 − p) . Hence

conditional mean of N - r is λ(1 − p) .

Page 4
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

4 The development factors are:

Development year
Accident 0 1 2 3 EP
Year

2003 3,340 3,750 4,270 4,400 4,800

2004 3,670 4,080 4,590 4,900
2005 3,690 4,290 5,050
2006 4,150 5,200
10,700 7,830 4,270
12,120 8,860 4,400
Development factors 1.1327 1.1315 1.0304
Cumulative DFs 1.3207 1.1660 1.0304

Ultimate loss ratio 4,400/4,800 = 0.9167

Estimated ultimate loss for each accident year:

Accident ULR EP UL Expected Claims to be

Year claims incurred

2003 0.9167 4,800 4,400 4,400 0

2004 0.9167 4,900 4,492 4,359 133
2005 0.9167 5,050 4,629 3,970 659
2006 0.9167 5,200 4,767 3,609 1,158

Revised ultimate losses are:

Accident Claims Claims to be Revised UL

Year incurred incurred

2003 4,400 0 4,400

2004 4,590 133 4,723
2005 4,290 659 4,949
2006 4,150 1,158 5,308
19,379
Claims to date 15,000
Reserve required 4,379

Page 5
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

5 (i) Let X be the size of the first claim, so that X has an exponential distribution
with parameter 1. Then for ruin to occur at time t we require
X > U + (1 + α) λt.

∞
P(X > U + (1 + α) λt) = ∫ e− x dx
U + (1+α ) λt

∞
= ⎡ −e − x ⎤
⎣ ⎦ U +(1+α )λt

= e-Ue-(1+α)λt.

[Note that it would be acceptable to quote the cumulative distribution function

for the exponential distribution from the tables rather than calculate the
integral]

(ii) Let T denote the time until the first claim. Then T has an exponential
distribution with parameter λ and

∞
P(Ruin at first claim) = ∫ P(Ruin at first claim⏐first claim is at t ) × fT (t )dt
0

∞
= ∫ e −U e −(1+α ) λt λe −λt dt
0

∞
−U
= ∫e λe−(2+α )λt dt
0

∞
⎡ λ ⎤
= ⎢ −e −U e −(2+α )λt ⎥
⎣ (2 + α)λ ⎦0

e −U
= .
2+α

Page 6
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

(iii) We require

e −U
< 0.01
2+α

i.e. e-U < 0.01 × (2 + α)

i.e. 100e-U < 2 + α.
i.e. 100e-U – 2 < α.

6 (i) We find a by solving:

∫ f ( x)dx = 1
0

2
2
−x
∫ axe dx = 1
0

2
⎡ a − x2 ⎤
⎢⎣ − 2 e ⎥⎦ = 1
0

− a −4
(e − 1) = 1
2

−2
a= −4
= 2.03731
e −1

(ii) We find the local maximum value of f(x) by differentiation:

2 2 2
f '( x) = ae − x − 2ax 2e− x = ae− x (1 − 2 x 2 )

and this derivative is zero when

2 x2 = 1
x = ±0.70711

and f(0.70711) = 0.8738.

Note that f(0) = 0 and f(2) = 0.07463 so that the maximum value on [0,2] is
0.8738 and so f ( x) < 1 as required.

Page 7
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

(iii) Take as our first point (2°0.7413, 0.4601) = (1.4826,0.4601)

Now f(1.4826) = 0.3353 which is less than 0.4601 so we reject this point as it
lies above the graph of f(x).

Take as our second point (2°0.3210, 0.6316) = (0.6420,0.6316)

Now f(0.6420) = 0.86615 > 0.6316 so this point lies below the graph of f(x)
and is therefore acceptable. Our random sample is therefore the x co-ordinate
= 0.6420.

(iv) The box 0 < x < 2, 0 < y < 1 has area 2, and the area under the curve of f(x) is
1 by definition. Therefore we expect half the points to be rejected as they lie
above f(x). Hence it will on average take 4 U(0,1) simulations to determine
one point using the acceptance-rejection method.

(v) It would be better to use the same simulated claims to evaluate the re-
insurance arrangements

This avoids the possibility that the apparent superiority of one arrangement is
in fact due to a favourable series of simulated claims

7 (i) Let θ = exp(-λ)

π=πP

⎡1 − θ θ 0 0⎤
⎢1 − θ 0 θ 0 ⎥⎥
P=⎢
⎢ 0 1− θ 0 θ⎥
⎢ ⎥
⎣ 0 0 1− θ θ⎦

π0 = (1 - θ) π0 + (1 - θ) π25
π25 = θ π0 + (1 - θ) π40
π40 = θ π25 + (1 - θ) π50
π50 = θ π40 + θ π50

π0 + π25 + π40 + π50 = 1

θ
π25 = π0 = kπ0
1− θ
θ
π40 = π25 = k2π0
1− θ
θ
π50 = π40 = k3π0
1− θ

and hence π0 + k π0 + k2 π0 + k3 π0 = 1

Page 8
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

1
hence π0 =
1+ k + k 2 + k3

hence the average premium paid is

⎛ 1 + 0.75k + 0.6k 2 + 0.5k 3 ⎞

500 × ⎜ ⎟⎟
⎜ 1+ k + k 2 + k3
⎝ ⎠

(ii) (a) θ/(1 - θ) = e-0.12/(1 – e-0.12) = 7.8433 Premium = £257.79

(b) θ/(1 - θ) = e-0.24/(1 – e-0.24) = 3.6866 Premium = £270.33
(c) θ/(1 - θ) = e-0.36/(1 – e-0.36) = 2.3077 Premium = £288.46

(iii) (a) to (b) λ increases by 100% but average premium paid increases only by
4.9%
(b) to (c) λ increases by 50% but average premium paid increases only by
6.7%

The no claims discount system is not effective at discriminating between good

and bad drivers.

8 (i) Let the individual loss amounts have distribution X. Then

100
−0.01333 x
E( X ) = ∫ 0.01333xe dx + 100 × P( X > 100)
0

100 ∞
100
= ⎡ − xe−0.01333 x ⎤ + ∫e
−0.01333 x
dx + 100 ∫ 0.01333e
−0.01333 x
dx
⎣ ⎦0
0 100

100 ∞
= −100e −1.333 + ⎡ −75e−0.01333 x ⎤ + 100 ⎡ −e−0.01333 x ⎤
⎣ ⎦0 ⎣ ⎦100

= −100e −1.333 − 75e−1.333 + 75 + 100e−1.333

= 55.2302

Hence E ( S ) = 50 × 55.2302 = 2761.5

Page 9
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

100
2 −0.01333 x
E( X 2 ) = ∫ 0.01333x e dx + 1002 P( X > 100)
0

100
100
= ⎡ − x 2e −0.01333 x ⎤ + ∫ 2 xe
−0.01333 x
dx + 1002 e−1.333
⎣ ⎦0
0

100 100
2 −1.333 ⎡ 2x ⎤ 2
= −100 e + ⎢− e−0.01333 x ⎥ + ∫ e−0.01333 x dx + 1002 e−1.333
⎣ 0.01333 ⎦0 0
0.01333

100
200 ⎡ 2 ⎤
=− e −1.333 + ⎢ − 2
e−0.01333 x ⎥
0.01333 ⎣ 0.01333 ⎦0

200 2 2
=− e −1.333 − 2
e−1.333 +
0.01333 0.01333 0.013332

= 4330.6

and so

Var ( S ) = 50 × 4330.6 = 216529 = (465.33) 2

(ii) (a) The normal distribution is N(2761.5, 465.332)

(b) The Log-Normal distribution has parameters μ and σ with

2
E ( S ) = eμ+σ /2

2 2 2
Var ( S ) = e 2μ+σ (eσ − 1) = E ( S ) 2 × (eσ − 1)

So substituting gives

2
216529 = 2761.52 × (eσ − 1)

2 216529
eσ = + 1 = 1.028394
2761.52

σ2 = log(1.028394) = 0.027998

σ = 0.167327

And now we can substitute for σ to give

Page 10
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

2761.5 = eμ+0.027998 / 2

μ = log(2761.5) − 0.027998 / 2 = 7.90953

(iii) Using the Normal distribution:

⎛ 3000 − 2761.5 ⎞
P ( N (2761.5, 465.332 ) > 3000) = P ⎜ N (0,1) > ⎟
⎝ 465.33 ⎠

= P( N (0,1) > 0.51) = 1 − 0.69497 = 0.30503

From tables.

Using the log-normal distribution,

P (log N (7.90953, 0.167327 2 ) > 3000) = P( N (7.90953, 0.167327 2 ) > log(3000))

log 3000 − 7.90953

= P( N (0,1) > ).
0.167327

= P( N (0,1) > 0.58) = 1 − 0.71904 = 0.28096

n n
μ yi e−μ
9 (i) The likelihood is ∏ f ( yi | μ ) = ∏
yi !
i =1 i =1

and hence the log-likelihood is

n n
log μ∑ yi − nμ − … = θ∑ yi − nb ( θ ) + terms not depending on θ
i =1 i =1

where θ = logμ
b(θ) = eθ

(ii) (a) The Pearson residual is

yi − yˆi
yˆi

(b) The Pearson residuals are skewed.

This makes it difficult to assess the fit of the model by eye.

Page 11
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

(iii) The conjugate prior has the same θ dependence as the likelihood, which is
{ }
proportional to exp yθ − eθ . Hence the conjugate prior is exp αθ − β eθ . { }
f ( θ | y1, y2 , …, yn ) ∝ f ( y1 , y2 , …, yn | θ ) f ( θ )
⎧⎪ n ⎫⎪
∝ exp ⎨θ∑ yi − neθ ⎬ exp αθ − β eθ { }
⎩⎪ i =1 ⎭⎪
⎧⎪ ⎛ n ⎞ ⎫⎪
∝ exp ⎨θ ⎜ α + ∑ yi ⎟ − ( β + n ) eθ ⎬
⎜ ⎟
⎪⎩ ⎝ i =1 ⎠ ⎭⎪

∂ log f
(iv) For the prior log f = αθ − β eθ and = α − β eθ . Hence
∂θ

⎡ ∂ log f ⎤ α
E⎢ ⎥ = E ⎡α − β eθ ⎤ =0, and so E ⎡eθ ⎤ = .
⎣ ∂θ ⎦ ⎣ ⎦ ⎣ ⎦ β

⎡ ∂ log f ⎤ ⎡ n ⎤
For the posterior E ⎢
⎣ ∂θ ⎥⎦
= E ⎢ α + ∑ i y − ( β + n ) e θ
⎥ , and hence
⎣⎢ i =1 ⎦⎥
n
α + ∑ yi
θ
E ⎡e | y1, y2 , …, yn ⎤ = i =1
.
⎣ ⎦ β+n

Note that eθ = μ , and the posterior estimate can be written as

n n

β α n
∑y i
α ∑y i
× + × i =1 = Z + (1 − Z ) i =1
β +n β β +n n β n

ie a combination of the prior estimate and the estimate from the data.

Page 12
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

10 (i) The characteristic equation is given by:

8 1 2 1 1
(1 - λ + λ ) = (1 - λ ) (1 - λ ) = 0
15 15 3 5

which has roots = 3 and 5. They are both greater than 1. Hence, subject to the
initial values having appropriate distributions, this implies (weak) stationarity.

(ii) (a)

Firstly, note that Cov( X t , Z t ) = 1 and Cov( X t , Z t −1 ) = 8 / 15 − 1 / 7 = 41 / 105

We need to generate 3 distinct equations linking γ 0 , γ 1 and γ 2

This can be done as follows:

(A)

γ 0 = Cov( X t , X t ) = Cov(1 + 8 / 15 X t −1 − 1 / 15 X t − 2 + Z t − 1 / 7 Z t −1 , X t )
= 8 / 15γ 1 − 1 / 15γ 2 + 1 − 1 / 7 × 41 / 105
= 8 / 15γ 1 − 1 / 15γ 2 + 694 / 735

(B)

γ 1 = Cov( X t , X t −1 ) = Cov(1 + 8 / 15 X t −1 − 1 / 15 X t − 2 + Z t − 1 / 7 Z t −1 , X t −1 )
= 8 / 15γ 0 − 1 / 15γ 1 − 1 / 7

(C)

γ 2 = Cov( X t , X t − 2 ) = Cov(1 + 8 / 15 X t −1 − 1 / 15 X t − 2 + Z t − 1 / 7 Z t −1 , X t − 2 )
= 8 / 15γ 1 − 1 / 15γ 0

Next stage is to solve these equations.

Substituting (C) into (A) gives

γ 0 = (8 / 15)γ 1 − (1 / 15)((8 / 15)γ 1 − (1 / 15)γ 0 ) + 694 / 735

so
(224 / 225)γ 0 = (112 / 225)γ 1 + 694 / 735
γ 0 = (1 / 2)γ 1 + 5205 / 5488

Page 13
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

Now substituting into (B) gives

γ 1 = 8 / 15((1 / 2)γ 1 + 5205 / 5488) − (1 / 15)γ 1 − 1 / 7

so
(4 / 5)γ 1 = 249 / 686
γ 1 = 1245 / 2744 = 0.4537

And

γ 0 = 1 / 2 × 0.4537 + 5205 / 5488 = 1.1753

γ 2 = 8 / 15 × 0.4537 − 1 / 15 × 1.1753 = 0.1636

Finally,

γ1 γ
ρ0 = 1, ρ1 = = 0.386, ρ2 = 2 = 0.139
γ0 γ0

8 1
(b) ρk = ρk −1 - ρk − 2 for k ≥ 2
15 15

We will show that the solution has the form:

1 1
ρk = A ( )k + B ( )k
3 5

Substituting the proposed solution into the recurrence relation gives

8 1 8 1 1 1 1 1
ρk −1 - ρk −2 = [ A ( )k-1 + B ( )k–1] - [ A ( )k–2 + B ( )k–2]
15 15 15 2 5 15 3 5

1 8 1 1 8 1
= A ( )k ( × 3 - × 9) + B ( )k ( ×5- × 25)
3 15 5 5 15 15

1 1
= A ( )k + B ( )k
3 5

= ρk

So the solution does have this form.

Page 14
Subject CT6 (Statistical Methods Core Technical) — September 2007 — Examiners’ Report

The values of A and B are fixed by ρ0 = 1, ρ1 = 0.386

∴A + B =1

1 1
A + B = 0.386
3 5

1 1
→ A + (1 – A ) = 0.386
3 5

A = 1.395

B = -0.395

1 1
∴ Pk = 1.395 ( )k – 0.395 ( )k
3 5

(iii) We require mean and variance of X t which must be normally distributed

since Z is normally distributed.

Variance is γ 0 = 1.1753 from (ii) (a)

8 1
E ( Xt ) = 1 + E ( Xt ) - E ( Xt )
15 15

15
∴ E ( Xt ) =
8

END OF EXAMINERS’ REPORT

Page 15