Faculty of Actuaries Institute of Actuaries

Subject CT6 — Statistical Methods

Core Technical

EXAMINERS’ REPORT
April 2008

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

June 2008

Comments

Comments for individual questions are given after each of the solutions that follow.

© Faculty of Actuaries
© Institute of Actuaries
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

1
Type Typical perils
Employers’ liability • Accidents caused by employer negligence
• Exposure to harmful substances
• Exposure to harmful working conditions
Motor 3rd party liability • Road traffic accidents
Public liability • Will relate to the type of policy
Product liability • Faulty design, manufacture or packaging of
product
• Incorrect or misleading instructions
Professional Indemnity • Wrong medical diagnosis, error in medical
operation etc.

Comment: Most of the candidates did very well here.

2 (i) P (μ < 270) = P( N (300, 202 ) < 270)

270 − 300
= P( N (0,1) < )
20
= P( N (0,1) < −1.5)
= 1 − 0.93319
= 0.06681

(ii)

(a) Using the result from page 28 of the tables, the posterior distribution of μ is
normal, with mean

10 × 270 300
(2
+ 2)
μ* = 50 20 = £281.54
10 1
( 2 + 2)
50 20

and variance

1
σ*2 = = 153.85 = 12.402
10 1
2
+ 2
50 20

So the posterior distribution of μ is N (281.54,12.402 ) .

Page 2
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

(b) The posterior probability required is given by:

270 − 281.54
P ( N (281.54,12.402 ) < 270) = P ( N (0,1) < )
12.4
= P ( N (0,1) < −0.931)
= 1 − (0.9 × 0.82381 + 0.1× 0.82639)
= 0.1759

Comment: The probability that the true mean is less than £270 has risen, as the
sample evidence suggests that the true mean is less than the mean of the prior
distribution. Nevertheless, the sample size is relatively small, and the variance of the
prior distribution is also small, so that a reasonable weight is still given to the prior
information.

Comment: Some candidates did not use tables for generating the posterior
parameters in (ii)(a). A few gave the correct interpretation in (ii)(b).

3 (i) M X (t ) = E (etX )
∞
λk
= ∑ e kt e−λ
k =0 k!
∞
(λ et ) k
= ∑ e −λ
k =0 k!

= e−λ eλe
t

= eλ ( e −1)
t

Hence

M X +Y (t ) = E (et ( X +Y ) )
= E (etX ) E (etY )
= M X (t ) M Y (t )
= eλ (e −1) eλ (e −1)
t t

= e 2λ (e −1)
t

which is the MGF of a Poisson distribution with parameter 2λ. Hence X + Y is

Poisson distributed.

(ii) Using the result above, aggregate claims on the portfolio over the year have a
Poisson distribution with parameter 18.

Let the initial capital be U. At the end of the year, the surplus will be
U + 20 – N where N is the number of claims.

Page 3
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

Now using the tables in the gold book, P(N <= 28) = 0.9897, and
P(N < = 29) = 0.9941.

So we need U to be large enough that ruin would only occur if there were 30
or more claims. So we need U + 20 – 29 > 0.

i.e. U > 9.

Comment: Many candidates struggled with (ii) here.

1 ⎧⎪ ( y − μ ) 2 ⎫⎪
4 (i) f(yi) = exp ⎨− i 2i ⎬
2πσ2 ⎩⎪ 2σ ⎭⎪

2 ( yi2 − 2 yi μi + μi2 )
logf = −½ log(2πσ ) −
2σ 2

yi μi − ½μi2 yi2
= − ½ log(2πσ2 ) − ½
σ2 σ2

which is the form of an exponential family of distributions.

(ii) The natural parameter is μi

θi = μi

b(θi) = ½μi2 = ½θi2

b′(θi ) = θi

b′′(θi ) = 1

Hence V(μi) = 1

(iii) The scaled deviance is

⎡ ( y − y )2 ( y − μˆ ) 2 ⎤
2 ∑ ⎢ − i 2i − ½ log(2πσ2 ) + i 2i + ½ log(2πσ 2 ) ⎥
⎣⎢ 2σ 2σ ⎦⎥

( yi − μˆ i )2
= ∑ σ2

Page 4
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

Hence the deviance residual is

( yi − μˆ i ) 2
yi − μˆ i
sign( yi − μˆ ) =
σ 2 σ
y − μˆ i y − μˆ
The Pearson residual is i = i
σV (μˆ i ) σ

Comment: This was a relatively easy question and many scored full marks in (i) and
(iii) here.

5 Multiply the claim payments with the corresponding inflation factors given below:

Development year

2004 1.16757 1.11197 1.05400 1.00000

2005 1.11197 1.05400 1.00000
2006 1.05400 1.00000
2007 1.00000

The resulting table is:

Development year

2004 478.70 905.14 227.66 79.00

2005 639.38 990.76 281.00
2006 857.96 1066.00
2007 1142.00

The inflation adjusted accumulated claim payments in mid 2007 are given below:

Development year

year 0 1 2 3
2004 478.70 1383.84 1611.50 1690.50
2005 639.38 1630.14 1911.14 2004.83
2006 857.96 1923.96 2248.66 2358.90
2007 1142.00 2853.75 3335.38 3498.88

Note only the values of the last row are needed for the answer.

The bolded values show the completed table using the basic chain ladder approach.

The development factors are 2.4989, 1.1688, 1.0490.

Page 5
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

For the answer we only need to work with the projected values at the last row as:

(2853.75-1142.00)*1.08+(3335.38-2853.75)*1.082+(3498.88-3335.38)*1.083
= 2616.43

2616.43 * 5000 = £13,082,150

Comment: Many candidates scored full marks here. Some missed the conversion of
the final figure from units to pounds (i.e. multiplying by £5000).

6 (i) E[S] = 65 × 1200 + 20 × 4500 = 168000

Var[S] = 65 × 2 × 12002 + 20 × 2 × 45002

= 187,200,000 + 810,000,000

= 997,200,000

(ii) (a) c = annual premium income = (1 + θ) E[S]

⎛ u + c − E[ S ] ⎞
P(S < u + c)  Φ ⎜
 ⎜ Var[ S ] ⎟⎟
⎝ ⎠

u + c − E[ S ]
= 1.96
Var[ S ]

u + θE[S] = 1.96 Var[ S ]

u = 1.96 Var[ S ] - θE[S]

= 61893.8 - 168000θ

(b) 61893.8 -168000θ = 0

θ = 0.3684

No initial reserve is required if θ ≥ 0.3684. i.e. when the premium ≥

£229,894.

Comment: There were some mixed answers for the second part of this question.
Numerical figures varied in (b) due to the rounding of the square root of the
Var(S).

Page 6
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

7 (i) Using the back shift operator it can be seen that

Yt = (1 + β1B) (1 + β4B4) et.

The invertibility conditions are then β1 < 1 and β4 < 1.

(ii) Since E(Yt) = 0, γ0 = E( yt2 ) = σ2 (1 + β12 + β24 + β12β24 ) = σ 2 (1 + β12 )(1 + β24 ).
Similarly it can be shown that

γ1 = E(YtYt-1) = σ2β1 (1 + β24 )

γ2 = 0

γ3 = σ2β1β4

γ4 = σ2β4 (1 + β12 )

γ5 = γ3

γk = 0, k > 5

So the ACF is

β1
ρ1 =
1 + β12

ρ2 = 0

β1β4
ρ3 = ρ5 =
(1 + β12 )(1 + β24 )

β4
ρ4 =
1 + β24

ρk = 0, k > 5

(iii) Since in general the ratio u ≤ 0.5 then we see that for our model
1+u 2
ρ1 < 0.5, ρ3 < 0.25, ρ4 < 0.5 and ρ5 < 0.25. These do not seem to be
satisfied by the sample ACF. So the model is not appropriate for such data.

Other observations like those listed below can suffice here:

• r(2) is not zero, and neither are r(6) and r(7).

Page 7
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

• r(3) is not close to r(5).

• r(1)r(4) = 0.43. This should be similar in value to both r(3) and r(5).
Whilst close to r(3) it isn't close to r(5).

Full marks for at least three correct statements.

Comment: There were some easy marks here. With the exception of part (iii), many
candidates did well but some dropped many points when the concept of auto-
correlation was not clear.

8 (i)

Level Prem. If. claim Prem. No claim Difference

0 800 600 600 400 400
1 800 600 400 400 600
2 600 400 400 400 200

(ii) The claims are exponentially distributed with parameter λ = 1/ μ = 1/ 600 and
so Pr(loss > u) = exp(-u λ).

Since
P(claim) = P(accident) P(claim|accident)

We derive that for a policyholder at Level 0

P(claim at 0%) = 0.2P(X > 400) = 0.2 exp(-400/600) = 0.102683

for Level 1
P(claim at 25%) = 0.2P(X > 600) = 0.2 exp(-600/600) = 0.07357589

and for Level 2

P(claim at 50%) = 0.2P(X > 200) = 0.2 exp(-200/600) = 0.1433063

(iii) The transition matrix will be

⎛ 0.1027 0.8973 0.0000 ⎞

⎜ ⎟
⎜ 0.0736 0.0000 0.9264 ⎟
⎜ 0.0000 0.1433 0.8567 ⎟
⎝ ⎠

π P = π and hence

π0 = 0.1027π0 + 0.0736π1
π1 = 0.8973π0 + 0.1433π2
π2 = 0.9264π1 + 0.8567π2

Page 8
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

The first and the last equations imply

π0 = 0.0736 / 0.8973π1 = 0.0820π1

and

π2 = 0.9264 / 0.1433π1 = 6.4748π1

From π0 + π1 + π2 = 1 we obtain π1 = 1/7.5568 = 0.1325

with π0 = 0.0820×0.1325 = 0.0109 and π2 = 1 - 0.1325 - 0.0109 = 0.8566.
The average premium is now:
800 π0 + 600 π1 + 400 π2 = 800 × 0.0109 + 600 × 0.1325 + 400 × 0.8566 =
430.85.

Comment: Generally, very good answers with many candidates scoring full marks.

9 (i) Strictly stationary processes have the property that the distribution of
(Xt+1, … Xt+k) is the same as that of (Xt+s+1, … Xt+s+k) for each t, s and k. For
the weakly stationary only the first two moments are needed to satisfy

E(Xt) = μ ∀t

and

cov(Xt, Xt+s) = γ(s) ∀t, s.

(ii) These two definitions coincide for the multivariate normal processes since the
normal distribution is characterised by the first two moments only.

(iii) In order to confirm that we need to calculate the eigenvalues of the parameter
matrix

⎛ 0.5 0.3 ⎞
A= ⎜ ⎟.
⎝ 0.1 0.8 ⎠

So we need to solve det(A - λI) = 0 which implies the solution of

(0.5 - λ) (0.8 - λ) – 0.03 = 0

0.37 − 1.3λ + λ 2 = 0

We see that this equation is satisfied for λ1 = 0.8791288 and λ2 = 0.4208712.

Since they are both smaller than 1, the process is stationary.

(iv) The parameter matrix here is Ac = A + cI, and the eigenvalues equation is now
det(A + cI -λI) = 0 or det(A – (λ - c)I) = 0.

Page 9
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

So the eigenvalues of Ac are λ1 + c and λ2 + c where λi are those of A.

Since λi are positive then the required values for c are such that λ1 + c < 1 and
λ2 + c < 1.
Hence 0 < c < 1 - λ1 = 0.1208712, since λ1 is the largest of the two.

Comment: This was not the easiest question. Some struggled with (ii), (iii) and (iv).
There were quite a few candidates who managed to avoid the calculation of the eigen
values of the matrix A by explicitly expressing each X_n and Y_n series as stationary
univariate AR(2) processes with some white noise terms.

10 (i) Let N denote the annual number of accidents. Then N ~ B(250,p) and (from
the tables) M N (t ) = ( pet + 1 − p) 250

If there is an accident, then the total cost of replacement wheels, X, has the
following distribution:

Number of wheels requiring replacement 0 1 2

Cost of replacement X £0 £100 £200
Probability 0.81 0.18 0.01
And M X (t ) = 0.01e200t + 0.18e100t + 0.81 .

So

M S (t ) = M N (log M X (t ))
= ( pelog M X (t ) + 1 − p) 250
= ( pM X (t ) + 1 − p ) 250
= ( p(0.01e200t + 0.18e100t + 0.81) + 1 − p) 250
250
⎛ pe 200t + 18 pe100t + 81 p ⎞
=⎜ +1− p ⎟
⎜ 100 ⎟
⎝ ⎠

(ii) E ( S ) = M S' (0)

249
⎛ pe 200t + 18 pe100t + 81 p ⎞
M S '(t ) = 250 × ⎜ +1− p ⎟ × (2 pe 200t + 18 pe100t )
⎜ 100 ⎟
⎝ ⎠
E ( S ) = M S '(0) = 250 ×1× 20 p = 5000 p

Page 10
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

E ( S 2 ) = M S ''(0)
248
⎛ pe 200t + 18 pe100t + 81 p ⎞
M S ''(t ) = 250 × 249 × ⎜ +1− p⎟ × (2 pe 200t + 18 pe100t ) 2
⎜ 100 ⎟
⎝ ⎠
249
⎛ pe200t + 18 pe100t + 81 p ⎞
+250 × ⎜ +1− p⎟ × (400 pe 200t + 1800 pe100t )
⎜ 100 ⎟
⎝ ⎠

M S ''(0) = 250 × 249 × (20 p ) 2 + 250 × 1× 2200 p = 24,900, 000 p 2 + 550, 000 p
Var ( S ) = E ( S 2 ) − E ( S ) 2 = 24,900, 000 p 2 + 550, 000 p − (5000 p ) 2 = 550, 000 p − 100, 000 p 2 .

Alternatively, we note that

E(N) = 250p and Var(N) = 250p (1 - p)

E(X) = 0.01 × 200 + 0.18 × 100 = 20
Var(X) = 40000 × 0.01 + 10000 × 0.18 – 20 × 20 = 1800

and

E(S) = E(N) E(X) = 250p × 20 = 5000p

Var(S) = E(N) Var(X) +Var(N) × E(X) × E(X)
Var(S) = 250p × 1800 + 250p(1 - p) × 20 × 20
= 450,000p + 100,000p (1 - p) = 550,000p - 100,000 p 2 .

(iii) Let W denote the total number of wheels needing replacement. Then
W~B(500,0.1p) and T = 100W

Hence

E(T) = 100E(W) = 100 × 500 × 0.1p = 5000p

and

Var (T ) = Var (100W ) = 1002Var (W ) = 1002 × 500 × 0.1 p × (1 − 0.1 p )

= 500, 000 p (1 − 0.1 p )

(iv) (a) If p = 0.05 then

E(S) = E(T) = 250.

Var ( S ) = 550,000 × 0.05 − 100,000 × 0.05 × 0.05 = 27,250 = 165.08 2

Var (T ) = 500,000 × 0.05 × 0.995 = 24,875 = 157.72 2

Page 11
 Subject CT6 (Statistical Methods Core Technical) — April 2008 — Examiners’ Report

(b) And so assuming both can be approximated by a normal distribution,

and allowing for a continuity correction

550 − 250
P ( S > 550) ≈ P( N (0,1) > ) = P( N (0,1) > 1.817)
165.08
= 1 − (0.7 × 0.96562 + 0.3 × 0.96485)
= 0.034611

550 − 250
P (T > 500) ≈ P ( N (0,1) > ) = P ( N (0,1) > 1.902)
157.72
= 1 − (0.8 × 0.97128 + 0.2 × 0.97193)
= 0.02859

(c) The two distributions have the same mean, but different variances – the
variance of S is slightly higher than that of T. This leads to a higher
probability for such a loss under S than under the approximation T.
Though the probabilities are both small in absolute terms, that for S is
20% higher than that for T. Effectively, fewer accidents are needed
under S to give a high loss, because each accident can lead to two
wheels being replaced, whereas under T only one wheel can be
damaged per accident.

Comment: This was a challenging question with many students scoring well here and
some trying to fudge the answers for (i).

END OF EXAMINERS’ REPORT

Page 12