Stochastic Modelling 2000-2004

Faculty of Actuaries
Institute of Actuaries
EXAMINATIONS
10 April 2000 (pm)
Subject 103 Stochastic Modelling
Time allowed: Three hours

INSTRUCTIONS TO THE CANDIDATE
1.
Write your surname in full, the initials of your other names and your
Candidates Number on the front of the answer booklet.
2.
Mark allocations are shown in brackets.
3.
Attempt all 11 questions, beginning your answer to each question on a

separate sheet.
Graph paper is not required for this paper.
AT THE END OF THE EXAMINATION
Hand in BOTH your answer booklet and this question paper.

In addition to this paper you should have available
Actuarial Tables and an electronic calculator.
103A2000
Let Xn = Y1 + Y2 + ... + Yn be a simple random walk with step distribution

P[Yj = 1] = p = 1 P[Yj = 1].
Derive an expression, for each > 0, for the two values of such that
Mn = e n+ X n is a martingale with respect to the natural filtration of Xn.
[4]
Let X and Y be jointly normal random variables with means X , Y , variances

2X , Y2 and correlation coefficient . Derive an expression for the values of
the coefficients and , given that the conditional expectation E[XY] is of the
form + Y.
[5]
The total number of claims received by an insurance company is described by

an inhomogeneous Poisson process with rate (t).
Write down the Kolmogorov forward equations for this process and show that,
as in the homogeneous case, the solution is of the form
P0j (s, t) =
where m(s, t) =
(m( s , t )) j e m ( s,t )
j!
z (x)dx.
t
s
[5]
According to a model used in econometrics, three economic time series, X, Y

and Z, are related to one another by the equations
Xn = Xn1 + X Zn1 + e1,n
Yn = Yn1 + Y Zn1 + e2,n
Zn =
Z Zn1 + e3,n
where each of X , Y and Z lies in the interval (1, +1) and the random
variables {ei,n : n 1, i = 1, 2, 3} may be assumed to be uncorrelated and to have
mean zero.
State, giving your reasons:
(a)
whether X, Y and Z are I(0) or I(1)
(b)
whether each of X, Y and Z individually satisfies the Markov property
(c)
whether the vector-valued process {(Xn , Yn , Zn) : n 1} satisfies the

Markov property
(d)
whether X and Y are cointegrated
1032
[5]
Let Bt be a standard Brownian motion, and let Ft = (Bs , 0 s t) be its natural

filtration.
(i)
Derive the conditional expectations E[ Bt2Fs ] and E[ Bt4Fs ] , where s t.

You may assume that the fourth moment of a random variable with
distribution N(0, 2) is 34.
(ii)
Hence construct a martingale out of Bt4 .
(i)
Apply the inverse transform method to generate an observation from

the density
f1(x) =
1
(1 + x )2
[4]
[3]
[Total 7]
(x > 0)
using a pseudo-random number u in the range 0 < u < 1. Explain how

this can be extended to generate an observation from the symmetrised
form of the same density
f2(x) =
(ii)
1
2(1 +x)2
(x R)
[3]
The Cauchy distribution has density function

f(x) =
( x + 2 )
2
(x R)
where is a positive parameter.

Show that
f(x) Cf2(x)
for all x R
2
( + 1). Hence devise a method based on Acceptance
Rejection sampling for generating observations from the Cauchy
distribution.
[4]
[Total 7]
as long as C
1033
PLEASE TURN OVER
(i)
(a)
Calculate the autocovariance function {k : k 0} and

autocorrelation function { k : k 0} of a first-order Moving
Average process
Xt = + et + 1 et1 ,
where {et : t 0} is a sequence of uncorrelated, zero-mean random
variables with common variance 2e .
(b)
(ii)
State the conditions on the values of the parameters such that

the process is invertible.
[5]
A sequence of observations x 1 , x 2 , ..., x n has sample variance $ 0 = 14.5,

sample lag-1 autocovariance $ 1 = 5.0. Show that there is more than one
first-order moving average process which can be fitted to these data,
but verify that only one of the fitted processes is invertible.
[4]
[Total 9]
The members of a health insurance scheme are classified as contributors or

beneficiaries; a member who is a contributor in one period becomes a
beneficiary in the next period if he or she becomes seriously ill, and this
happens with probability 0.1. The probability of a serious illness continuing
into the next period is 0.2. The rules of the scheme specify that any member
who is a beneficiary for three successive periods must become a contributor for
the next period; if the illness still persists the member may thereafter revert
to being a beneficiary.
(i)
(a)
Construct a discrete time Markov chain to model this health

scheme, introducing if necessary various classes of beneficiaries
and contributors (a five state model is suggested).
(b)
Draw the transition graph.
(c)
Write down the transition matrix of the chain.
[6]
(ii)
Explain whether the above Markov chain is irreducible, periodic or

both.
[2]
(iii)
(a)
Calculate the stationary probability distribution of the chain.
(b)
Determine the proportion of beneficiaries among the

membership in the stationary rgime.
(iv)
1034
[4]
Let b be the average gross payout per beneficiary and c the average
gross payout per contributor per period; this means that the nett
payments are b f and c f respectively, where f is the membership fee
per period (assumed to be uniform over members and over time).
(a)
Explain how b, c and f should be related if the scheme is to be

viable.
(b)
Calculate the average profit per period per member in the

stationary rgime if b = 600, c = 150 and f = 300.
[3]
[Total 15]
A stationary second-order autoregressive process X, which may be assumed to

be in equilibrium at time 0, is defined by
Xt = + 1(Xt1 ) + 2 (Xt2 ) + et ,
where {et : t 1} is a sequence of independent, zero-mean Normal random
variables, each with variance 2e .
(i)
(ii)
(a)
Obtain an equation for 1 in terms of 0 and 2 by substituting for

Xt in the equation 1 = Cov(Xt , Xt1).
(b)
Derive similar equations for 2 and 0.
(c)
State the autocorrelation function k of X for k = 0, 1, 2.
[5]
Suppose that the equations derived in (i) for 1 and 2 are used as the
basis of an estimation procedure: estimates $ 1 and $ 2 are defined to be
the solutions of those equations when is replaced by a suitablydefined sample autocorrelation function r.
Solve these equations.
10
(i)
[3]
[Total 8]
(a)
Define standard Brownian motion Bt , t 0 and give its transition

probability density.
(b)
Write down the transition probability density of general

Brownian motion Wt = Bt + t.
[4]
Let St defined represent a share price at time t.

(ii)
Solve the stochastic differential equation

dSt = Stdt + StdBt .
[5]
(iii)
Calculate, given the parameters = 25% p.a., = 20% on an annual

basis, the probability that the share price will exceed 45 in four months
time given that its current price is 38.
[4]
(iv)
Calculate the probability that the share price will exceed 45 at any
stage during the next four months given that its current value is 38.
[You may use the formula
P[max( Bs + s) > y ] = G
0s t
FG t y IJ + e
H t K
2 y
FG y t IJ ,
H t K
where y 0 and G denotes the normalised Gaussian probability

distribution function.]
[4]
[Total 17]
1035
PLEASE TURN OVER
11
Patients arriving at the Accident and Emergency department (state A) wait for
an average of one hour before being classified by a junior doctor as requiring
in-patient treatment (I), out-patient treatment (O) or further investigation (F).
Only one new arrival in ten is classified as an in-patient, five in ten as outpatients.
If needed, further investigation takes an average of 3 hours, after which 50% of
cases are discharged (D), 25% are sent to receive out-patient treatment and
25% admitted as in-patients.
Out-patient treatment takes an average of 2 hours to complete, in-patient
treatment an average of 60 hours. Both result in discharge.
It is suggested that a time-homogeneous Markov process with states A, F, I, O
and D could be used to model the progress of patients through the system,
with the ultimate aim of reducing the average time spent in the hospital.
(i)
Write down the matrix of transition rates, {ij : i, j = A, F, I, O, D}, of

such a model.
[2]
(ii)
Calculate the proportion of patients who eventually receive in-patient

treatment.
[1]
(iii)
Derive expressions for the probability that a patient arriving at time

t = 0 is:
(iv)
(a)
yet to be classified by the junior doctor at time t, and
(b)
undergoing further investigation at time t
[4]
Let mi denote the expectation of the time until discharge for a patient
currently in state i.
(a)
Explain in words why mi satisfies the following equation:

1
+
i
mi =
j {i , D}
ij
i
mj
where i = ij .
j
(b)
(v)
Hence calculate the expectation of the total time until discharge

for a newly-arrived patient.
[4]
State the distribution of the time spent in each state visited according
to this model.
[1]
The average times listed above may be assumed to be the sample mean waiting
times derived from tracking a large sample of patients through the system.
(vi)
1036
Describe briefly what additional feature of the data might be used to

check that this simple model matches the situation being modelled. [2]
(vii)
1037
The hospital management committee believes that replacing the junior

doctor with a more senior doctor will save resources by reducing the
proportion of cases sent for further investigation. Alternatively, the
same resources could go towards reducing out-patient treatment time.
(a)
Outline briefly the calculations that would need to be performed

to compare the options.
(b)
Discuss whether the current model is suitable as a basis for

making decisions of this nature.
[4]
[Total 18]
EXAMINATIONS
April 2000
EXAMINERS REPORT
Subject 103 (Stochastic Modelling) April 2000 Examiners Report
E[Mn+1Fn] = E[ e - l( n +1 )+ gX n +1Fn ]
= e - l(n +1) E[ e g ( X n +Yn +1 )Fn ] = e - l ( n+1) eg X n E[ eg Yn+1Fn ]
= e - l M n E[ e gYn +1 ] = e - l M n ( pe g + (1 - p ) e - g ).
Hence the condition for martingale:
peg + (1 - p)e-g = el .
Multiply by eg and solve quadratic equation in unknown eg:
eg =
e l e 2l - 4 p(1 - p )
.
2p
Assume E[XY] = a + bY and determine a, b using:

(i)
orthogonality condition E{(X - E[XY])Y} = 0.
(ii)
E{E[XY]} = E[X].
(i) gives E[XY] - aE[Y] -bE[Y2] = 0;

Since the correlation coefficient r is
r=
E[ XY ] - E[ X ]E[Y ]
,
s X sY
we have E[XY] = rsX sY + mX mY and (i) yields

amY + b( sY2 + mY2 ) = rsX sY + mX mY .
(ii) gives a + bmY = mX .

Solve the two simultaneous equations to get
b=
Page 2
HI X
,
IY
a = mX -
rs X
mY
sY
-0
l( t )
l( t )
F -l(t)
G
A(t) = G
GG
H
l( t )
l( t )
- l( t )
l( t )
- l(t )
L.
I
JJ
.
OJ
J
OK
Forward equations:
P(s, t) = P(s, t) A(t),

t
t s.
P00(s, t) = -l(t) P00(s, t) and P00(s, s) = 1 imply that P00(s, t) = exp(t

e-m(s,t).
t
s
l(u)du) =
P0j(s, t) = l(t) P0, j-1(s, t) - l(t) P0j(s, t) with initial

t
condition P0j(s, s) = 0.
For j > 0, we have
Verify that the form of P0j(s, t) given in the question satisfies this equation:
LHS = (jm(s, t)j-1 - m(s, t)j)
RHS = l(t)
m( s, t) j e -m( s ,t )
m( s, t ) j -1 e -m( s,t )
- l( t )
j!
( j - 1)!
The observation that
e -m( s ,t )
m( s, t),
j ! t
m( s, t ) = l(t) is sufficient to finish the verification.

t
(a)
Z is stationary, i.e. I(0), as it is a first-order autoregression;

X is not stationary but X is just a linear combination of Z and e1 , so is
stationary; this implies that X is I(1). The same goes for Y.
(b)
Z satisfies the Markov property on its own; X and Y do not, since they
depend on values of Z.
(c)
(X, Y, Z) is Markov; indeed, it is a vector autoregression.
(d)
X and Y are not cointegrated. Although both are I(1), any linear
combination W =aX + bY satisfies Wn = Wn-1 + qW Zn-1 + e3,n which does not
define a stationary process.
Page 3
(i)
E[ Bt2Fs ] = E[(Bt - Bs + Bs)2Fs]
= E[(Bt - Bs)2 + 2(Bt -Bs) Bs + Bs2Fs ]

= E[(Bt - Bs)2Fs] + 2BsE[Bt - BsFs] + Bs2 ,
by the property of conditional expectations which allows one to take out
what is known. Moreover, by independence of the increments, the above is
E[(Bt - Bs)2] + Bs2 = t - s + Bs2 .
Similarly,
E[ Bt4Fs ] = E[(Bt - Bs + Bs)4Fs]
= E[(Bt - Bs)4 + 4(Bt - Bs)3 + 6(Bt - Bs)2 Bs2 + 4(Bt - Bs) Bs Bs3 + Bs4Fs ]
= E[(Bt - Bs)4] + 6Bs2 E[(Bt - Bs)2] + Bs4 ,
where we used the independence of increments property as well as the fact
that moments of odd order of N(0, s2) vanish. Finally
E[ Bt4Fs ] = Bs4 + 6(t - s) Bs2 + 3(t - s)2.
(ii)
From above
E[ Bt4 - 6tBt2Fs ] = Bs4 + 6(t - s) Bs2 + 3(t - s)2 - 6t(t - s + Bs2 )
= Bs4 + 6sBs2 + 3(t - s)2 - 6t(t - s) = Bs4 - 6sBs2 + 3(s2 - t2)
\ Bt4 - 6tBt2 + 3t 2 is a martingale.
(i)
u = F1(x) =
x
u
is solved by x = F1-1 ( u ) =
.
1+ x
1-u
For the symmetrised version, the simplest thing is to multiply x by a

variable y which takes 1 depending on whether another pseudo-random
uniform number v is in the range (0, 0.5) or (0.5, 1).
(ii)
Page 4
2q(1 + x ) 2
By symmetry we only need consider x > 0, so we find maxx>0
.
p( q 2 + x 2 )
Differentiating the logarithm of this fraction and setting equal to 0, we get
2x
2
, with solution x = q2. Substituting this value in, we obtain
= 2
1+ x
q + x2
the required value of C.

f ( x q )
, which we observe is less than or equal to 1 everywhere.
Cf2 ( x )
The method of Acceptance-Rejection sampling goes as follows: use (i) to
generate a variable y from density f2. We accept y as a valid observation
from f(xq) with probability g(y), otherwise reject it. (Generate a uniform
variable u, and reject if u > g(y).) If we reject it, go back and generate
another y from f2 , and continue to do the same until eventual acceptance.
Let g(x) =
(i)
(a)
g0 = Var(et + b1 et-1) = (1 + b12 ) s e2 and g1 = Cov(et + b1 et-1 , et-1+ b1et-2)

= b1 s 2e , with gk = 0 for k > 1.
This gives r0 = 1, r1 = b1 / (1 + >12 ), rk = 0 otherwise.
(b)
Invertibility requires that b1 < 1, so that the sum Xt - b1 Xt-1 +

b12 X t-2 + ... converges. m and se are irrelevant.
We need to solve (1 + b12 ) s e2 = 14.5, b1 I 2e = 5.0. Eliminating s 2e , we
(ii)
have 1 + b12 = 2.9b1 , or b1 = (2.9 2.92 - 4 ) = 2.5 or 0.4.

b1 = 2.5 corresponds to I 2e = 2, whereas b1 = 0.4 corresponds to I 2e =
12.5.
For invertibility, solve 1 + b1z = 0. In the first case, z = -0.4 (no
good); in the second, z = -2.5 (OK).
(i)
(a)
States:
C: healthy contributor
C : contributor but ill
B1 , B2 , B3 : beneficiary, with index giving duration of illness
(b)
Transition
graph:
0.9
0.1
0.8
0.8
B1
0.8
B2
0.2
0.8
0.2
B3
0.2
C
0.2
Page 5

(c)
Transition matrix (states ordered C, C , B1 , B2 , B3 ):
F 0.9
GG 0.8
GG 0.8
GH 00..88
P=
01
.
0.2
0
0
0
0
0.2
0
0.2
I
0J
J
0J
0.2J
J
0K
0
(ii)
The chain is irreducible by inspection: every state is accessible from every

other state. State C is clearly aperiodic because of the one-step loop from C
to C; because of irreducibility, every other state must be aperiodic too.
(iii)
(a)
p = pP reads
pc = 0.9pc + 0.8(pc + p1 + p2 + p3)
pc = 0.2p3
p1 = 0.1pc + 0.2pc
p2 = 0.2p1
p3 = 0.2p2 .
Discard first equation and choose pc as working variable:
p3 =
1
pc = 5pc
0.2
p2 =
1
p3 = 5p3 = 25pc
0.2
\ p = pc(1248, 1, 125, 25, 5)

p1 =
1
p2 = 5p2 = 125pc
0.2
pc =
1
0.2
p1 pc = 10p1 - 2pc = 1248pc .
01
.
01
.
Find pc by normalisation: pc(1248 + 1 + 125 + 25 + 5) = 1,

\ pc =
(b)
Page 6
1
.
1404
Proportion of beneficiaries:
125 + 25 + 5
= 11.04%.
1404

(iv)
(a)
Average profit per period per member in stationary rgime is

Z = (f - c)
=f-c
FG 1248 + 1IJ + (f - b) FG 125 + 25 + 5IJ

H 1404 K
H 1404 K
1249
155
-b
.
1404
1404
For Z > 0 you need f > c

(b)
With the given data

Z = 300 -
(i)
1249
155
+b
.
1404
1404
150 1249 + 600 155

= 100.32.
1404
(a)
g1 = Cov(Xt , Xt-1) = Cov(a1Xt-1 + a2Xt-2 + et , Xt-1) = a1g0 + a2g1+ 0, since

et is independent of Xt-1.
(b)
Similarly g2 = a1g1 + a2g0 and g0 = a1g1 + a2g2 + Cov(Xt , et). A further

application of the same technique gives Cov(Xt , et) = s 2e .
Thus g1 =
(c)
(ii)
rk is found by the relation rk = gk / g0.
We have = 1 = r1 (1 - a 2 ) and a 2 +
= 1 =
10
(i)
F
GH
(a)
I
JK
a1
a12
g0 and g2 = a 2 +
g0 .
1 - a2
1 - a2
r1 (1 - r2 )
,
1 - r12
= 2 =
a 12
= r2 , which are solved by
1 - a 2
r2 - r12
.
1 - r12
Bt defined by following properties:
Independent increments: Bt - Bs independent of Ba , 0 a s

whenever s t.
Stationary Gaussian increments: Bt - Bs ~ N(0, t - s).
Continuous sample paths: t Bt continuous.
Page 7

Transition density to go from x at time s to y at time t:
1
gt-s(y - x) =
(b)
2p( t - s )
e -( y -x )
/ 2( t -s )
{Ws = x, Wt = y} = {sBs + ms = x, sBt + mt = y} = {Bs =

y - mt
}. Hence transition density of W is
s
1
gt-s
I
(ii)
FG y - x - m(t - s) IJ .
H
K
s
By Its lemma
d(log St) =
I
JK
F
GH
1
1
1
dSt +
- 2 ( dSt ) 2
2
St
St
= dt + sdBt -
I2
dt.
2
Hence
F
GH
log St = log S0 + m -
s2
2
I t + sB ,
JK
t
and finally
F m - s I t + sB
G 2 JK
S eH
.
2
St =
(iii)
LM F s I t > log b OP
a PQ
MN GH 2 JK
L 1 F b F s I tI OP
= P MB > G log - G m MN s H a H 2 JK JK PQ
F log b - F m - s I t I
GG a GH 2 JK JJ
= 1-G
JJ .
GG
s t
K
H
P[St > bS0 = a] = P sBt + m -
Page 8
x - ms
, Bt =
s

Here a = 38, b = 45, m = 0.25, s = 0.2, t =
1
3
year.
So, above quantity is

1 - G(0.800) = 1 - 0.7881 = 0.2119.
(iv)
LM
MN
F
GH
F
GH
OP
I I
JK JK
PQ
F F m - s I t + log b I
G G 2 JK
aJ
bU G H
JJ .
log V G aW G
s t
JK
GH
P max Ss bS0 = a = P max Bs + m 0 s t
0st
F F m - s I t - log b I
GG GH 2 JK
aJ
R 2m - s
J
=G
+ expS
JJ T s
GG
s t
K
H
2
s2 s
b
1
log
s
a
2 s
2
The first term is 0.2119 by (iii).
F bI
The second term is the product of G J
H aK
2m - s2
s2
= 6.9893 with G(-2.128)
= 1 - G(2.128) = 1 - 0.9833.
So the result is finally 0.2119 x 0.0167 = 0.3286.
11
(i)
The generator matrix of the process would be
F -1
GG 0
GG 0
GH 00
I
JJ
JJ .
J
0K
0.4
01
.
0.5
1
3
1
12
1
- 60
1
12
1
6
1
60
1
2
0
0
0
0
0
0
- 12
0
1
4 1
1
+
= .
10 10 4
5
(ii)
The probability of ever visiting state I is
(iii)
(a)
d
pAA ( t ) = -pAA(t), which has solution pAA(t) = e-t.
dt
(b)
Similarly,
d
1
pAF ( t ) = - pAF (t) + 0.4pAA(t), so that
dt
3
d t/3
{e pAF(t)} = 0.4et/3 pAA(t) = 0.4e-2t/3,
dt
giving pAF (t) = e-t/3 0.6(1 - e-2t/3).
Page 9

(iv)
(a)
The equation arises as follows: when the process arrives in state i the
subsequent holding time has mean l-1
i , after which the process
jumps to a different state, choosing state j with probability pij = sij / li
(independent of the length of the holding time). The total time to
reach state D is therefore the time until the first jump plus the time
from arriving in the new state until hitting D (unless the new state is
D).
(b)
We have mI = 60, mO = 2, mF = 3 +
1
1
60 +
2 = 18.5,
4
4
mA = 1 + 0.1 60 + 0.5 2 + 0.4 18.5 = 15.4 hours.
(v)
The time-homogeneous Markov model has exponential holding times, so the

distribution is completely determined by the expectation.
(vi)
A simple check on whether the Markov model fits the data is therefore to
verify that the distributions of holding times are at least roughly
exponential, and a simple way of doing that is to compare sample standard
deviations with sample means. More detailed comparisons might be
possible, depending on the size of the data set.
(vii)
(a)
Calculations required in the first case would include working out the
expected duration of stay if the change were implemented, which
involves solving the equations in (iv) again. For the second situation,
just replace mO in the original calculation. New parameter values
will need to be guessed. Whichever model comes out better should be
compared with the initial situation, to determine whether the
improvement was worth the additional resources.
(b)
Model suitability: on the one hand the required decision is couched in

terms of expectations, which lend themselves well to Markov process
treatment. On the other, the fundamental problem in the system is
queue length, which can never be successfully modelled by a process
which tracks only a single individual at a time. (A network of
queuing processes would be a much better model.)
Page 10
EXAMINATIONS
11 September 2000 (pm)

1.
2.
3.

separate sheet.

In addition to this paper you should have available Actuarial
Tables and an electronic calculator.
103S2000
There are N individuals in a population, some of whom have a certain infection

that spreads as follows. Contacts between two members of this population occur
in accordance with a Poisson process having rate . When a contact occurs, it is
equally likely to involve any of the P =
N
e2j
pairs of individuals in the population.
If a contact involves an infected and a non-infected individual, then with

probability p the noninfected individual becomes infected. Once infected, an
individual remains infected throughout. Let X(t) denote the number of infected
members of the population at time t.
(i)
State whether X(t), t > 0 is a continuous-time Markov jump process. If so,

write down its state space and transition rates; if not, explain how it can
be expressed in terms of a different process which is Markov.
[2]
(ii)
Derive an expression for the expected time until all members are infected,
starting from a single infected individual
[2]
[Total 4]
During a long motorway journey a child amuses himself by noting down, at the
end of each minute, the lane in which the car is travelling. The motorway has
three lanes and the journey lasts N minutes.
(i)
Describe how to fit a three-state time-homogeneous Markov chain model

to the data, writing down formulae for the estimates of the transition
probabilities.
[2]
(ii)
Describe one test which could be applied to determine whether the process
possesses the Markov property. [2]
[Total 4]
A standard Ornstein-Uhlenbeck process may be defined as a stationary zeromean Gaussian process {Ut : t R} with autocovariance function given by
Cov(Ut , Us) =
2 t s
e
2
(s, t R).
(i)
Show that the process {Un : n = 1, 2, ...}, obtained by observing U only at

integer times, is a first-order autoregression.
[2]
(ii)
Derive expressions for the parameters and 2 of the autoregression in

terms of and 2.
[2]
[Total 4]
1032
(i)
State Its lemma as it applies to a stochastic process {Xt : t 0} and a

function f(Xt) which is not explicitly dependent on t.
[2]
(ii)
Apply Itos Lemma with f(x) = x4 to calculate the stochastic differential

d( Bt4 ), where Bt is standard Brownian motion.
[3]
(iii)
Hence express the It integral

integral involving Bs.
z0
Bs3 dBs in terms of Bt and of an ordinary

[2]
[Total 7]
Consider a homogeneous Markov chain with state space S = {1, 2, 3} and

transition matrix
P=
F
G
G
H
I
0 J .
J
0K
(i)
Calculate the 3-step transition matrix.
(ii)
Calculate, for each of the following initial conditions, the probability that
the chain will be in state 3 when it is observed at time n = 3 given that:
(iii)
1033
[2]
(a)
the chain is in state 1 at time zero
(b)
the chain is in state 1 at time zero and in state 2 at time 1
(c)
the probabilities of being in states 1, 2, and 3 at time zero are

8
, 9 and 31
respectively
given by 14
31 31
[4]
How would your answers to (a), (b) and (c) change if the time of the
observation were n = 300 instead of n = 3?
[2]
[Total 8]
PLEASE TURN OVER
The daily closing price of a share is observed every trading day for a year,
yielding a sequence of values {s1 , ..., sn}. A model is required for the purposes of
predicting future variability of the share price. The model suggested is a
Brownian one.
(i)
Explain briefly, on purely theoretical grounds, which of the two models

I:
II:
St = + t + Bt
log(St) = + t + Bt
you would expect to provide a better fit.

(ii)
(iii)
Refer to Figures 1 and 2 below.

(a)
Explain briefly whether your chosen model appears to provide a

good fit to the data.
(b)
State one of the tests you could carry out on the data to ascertain
whether the model fits adequately.
[3]
(a)
Describe how a Lvy process model differs from a Brownian model.
(b)
Outline the difficulties would you encounter in practice if you were

fitting a Lvy process model to the data provided.
[3]
[Total 7]
Figure 1: the share price Sn
Figure 2: log-transformed share price log Sn
1034
[1]
A client wishes to model the behaviour of a stochastic process {Xt : t 0} which

represents the average annual return for a particular class of asset. After a
number of observations the client has determined that Corr(Xt , Xt1) = 0.7 and
Corr(Xt , Xt2) = 0.5. He thinks that one of the two models
I:
Xt = + 0.7(Xt1 ) + 0.5(Xt2 ) + et
II:
Xt = + et + 0.7et1 + 0.5et2
will be best, but cannot decide which. He has simulated both processes from time
t = 1 to time t = 200, but has not obtained the results he expected, so is seeking
your advice.
(i)
(a)
Outline a suitable method of simulating a second-order

autoregression, assuming you have access to a reliable stream
{uk : k 0} of pseudo-random numbers uniformly distributed over
the range [0, 1].
(b)
Explain why might it be desirable to ensure that the stream {uk}

can be re-used if necessary.
[4]
(ii)
State why neither of the suggested models is suitable.
(iii)
(a)
[1]
Derive the lag-1 and lag-2 autocorrelations, 1 and 2, of a secondorder autoregressive process
Xt = + 1 (Xt1 ) + 2 (Xt2 ) + et .
(b)
1035
Find values of the parameters 1 and 2 which would provide a

suitable AR(2) model for {Xt : t = 0, 1, 2, }.
[5]
[Total 10]
PLEASE TURN OVER
Consider a time-homogeneous Markov jump process {X(t) : t 0} with two states

denoted by 0, 1, and transition rates 0,1 = , 1,0 = .
(i)
State Kolmogorovs forward equation for the probability P0,0(t) that X is in

state 0 at time t, given that it starts in state 0.
[1]
(ii)
Show that P0,0(t) =
(iii)
Let Ot denote the total amount of time spent in state 0 up until time t,
e ( + )t .
+
+ +
[3]
1 if X s = 0
I sds , where Is =
. Derive,
0 if X s 0
using the result in part (ii), an expression for E[OtX(0) = 0], the expected
occupation time in state 0 by time t for the two-state continuous-time
Markov chain starting in state 0.
[2]
which may be expressed as Ot =
z0
(iv)
Write down the expected occupation time in state 1 by time t for the twostate continuous-time Markov chain starting in state 0.
[1]
(v)
A health insurance scheme labels members as healthy (state 0) or

unhealthy (state 1) at any time. When in state 0, members pay
contributions at rate ; when in state 1 they receive benefit at rate .
Expenses amount to a constant per member per unit time.
(a)
Explain how the above model can be used to calculate in terms

of and .
(b)
State the assumptions which you make in applying the model.
(c)
Discuss whether they are likely to be satisfied in practice.

[4]
[Total 11]
Suppose that the evolution of the price of an asset follows the lognormal model
log(St) = Yt = y + t + Bt where Bt denotes the standard Brownian motion and
is a negative drift. The asset will be liquidated at the stopping time
T = inf{t : Yt = a} when its value reduces to ea, where a is some number less than
y. Consider now the exponential Vt = exp(uYt c(u)t).
(i)
Derive the condition on c(u) under which {Vt : t 0} is a martingale.
[3]
(ii)
State the optional stopping theorem and explain how it is used.
[3]
(iii)
Derive the moment generating function f(y, v) = E[ e vTu Y0 = y] of the

bankruptcy time for positive v by applying the optional stopping theorem
[5]
to the martingale Vt .
[Total 11]
1036
10
Consider a survival model with two states alive (A) and dead (D), with timedependent transition rate from A to D equal to (t) = t. The time parameter, t,
represents the age of the individual under consideration.
(i)
Calculate the transition probability PAA(s, t), defined by

PAA(s, t) = P(X(t) = AX(s) = A).
(ii)
[2]
Show, by making use of the formula

E(X) =
z0
P[X x] dx
for a positive random variable X, that the expected future lifespan of an

individual aged s is
E[Rs] =
1 1 G(s )
g(s )
where G is the standard Gaussian probability distribution function and g

is its density.
[4]
(iii)
It is desired to calibrate the above model so that the expected future

lifespan of an individual aged 70 is 6 years. Derive an approximation to
the corresponding value of , using the double inequality
1
1 1 G( x ) 1
.
x x3
g( x )
x
(iv)
[5]
A company wishes to test the validity of the above model. They assume
that the true force of mortality from age 70 onwards is of the form
(t) = a + bt and intend to test whether a = 0. The testing method will be
to simulate one sample of size 1000 when a = 0 and another when a 0,
then to see which most resembles the data which the company has
collected.
Explain how to simulate a value from the proposed distribution, for
arbitrary values of a and b.
[5]
[Total 16]
1037
PLEASE TURN OVER
11
The movements of a consumer price index are to be subjected to time series

analysis with the aim of forecasting future behaviour. The index is calculated
monthly.
(i)
Explain whether you would expect to fit a model which included (a) a
trend term, (b) a seasonal effect.
[3]
The values {xt : 1 t n} are the residuals which remain once any trend or
seasonal variations have been removed. An ARIMA(1, 1, 1) model is to be fitted
to the {xt}.
(ii)
(iii)
(a)
Assuming the ARIMA(1, 1, 1) model is correct, write down an

equation for Xn+1 in terms of the white noise process
{et: 1 t n + 1} and the observations {xt: 1 t n}.
(b)
State the parameters of the model.
[2]
The Box-Jenkins procedure defines the k-step-ahead forecast for X to be
x n ( k ) = E(Xn+kxn , xn1 , ..., x1).
(iv)
(a)
Derive the 1-step-ahead and 2-step-ahead forecasts for X for the

ARIMA(1, 1, 1) model, assuming that the values of the parameters
and the value of e0 are known exactly.
(b)
Evaluate the prediction variance Var(Xn+1 x n (1)) , again assuming

that the values of the parameters are known.
[5]
The most elementary form of the technique known as exponential

smoothing produces at time n a 1-step-ahead forecast x n* defined by
x n* = xn + ( x n* 1 x n ),
for some (0, 1) which may be chosen by the user.

Show that, for particular values of the autoregressive and moving average
parameters, the Box-Jenkins forecasts above coincide with the forecasts
produced by exponential smoothing.
[2]
(v)
(vi)
1038
(a)
State whether the ARIMA(1, 1, 1) model is I(0), I(1) or neither.
(b)
Discuss whether there is a difference between an I(0) model and an

I(1) model in terms of the conditional distribution of Xn+k given {xt :
1 t n} for large values of k. [3]
It is suggested that a salaries index might be cointegrated with the

consumer price index.
(a)
Explain what is meant by the suggestion.
(b)
Comment on whether it is a reasonable suggestion.
[3]
[Total 18]
EXAMINATIONS
September 2000
EXAMINERS REPORT
Subject 103 (Stochastic Modelling) September 2000 Examiners Report
(i)
This is clearly a Markovian birth process. The state space is {0, 1, ..., N}.
Given that we have m infected and N m healthy individuals, the
number of dangerous pair contacts is m(N m). Thus, the rate
m( N m)
m,m+1 = p
.
P
(ii)
The expected total infection time is the sum of the reciprocal rates
N 1
P
1
.
m
=
1
m( N m)
p
1
1 1
1
=
+
, this time may also be expressed as

m( N m)
N m N m
N 1 1
N 1
.)
m
=1 m
p
(Since
(i)
Let Nij denote the number of minutes when the car was in lane i at the
start of the minute and in lane j at the end. The estimate of the
transition probability pij is Nij / Ni+ , where Ni+ = j Nij .
(ii)
The problem here is the alternative hypothesis. It would be possible to

test whether the distribution of Xn+1 conditional on Xn = i was really
independent of Xn1. Or one might test, using a standard goodness-of-fit
test, whether the distribution of the number of consecutive minutes spent
in lane i really was geometrically distributed with parameter determined
by i.
(i)
Since {Un} is Gaussian and stationary, it is determined uniquely by its

mean and autocovariance functions. For k > 0 we have k = Cov(Un , Unk)
=
(ii)
2 k
2
e , so that the ACF is k = ek and the variance 0 =
.
2
2
Compare this with the corresponding values for an AR(1): 0 =
2
and
1 2
k = k for k > 0.
The two are seen to match as long as = e and 2 = (1 e2)
Page 2
2
.
2
(i)
If X satisfies dXt = Yt dBt + Zt dt, then f(Xt) satisfies

d[f(Xt)] = f (Xt) Yt dBt + { f (Xt) Zt + f (Xt) Yt2 } dt.
(ii)
(x4) = 4x3 , (x4) = 12x2.

d( Bt4 ) = 4 Bt3 dBt +
(iii)
. 12Bt2dt = 4 Bt3 dBt + 6 Bt2dt.
d( Bs4 ) = 4 t0 Bs3 dBs + 6 t0 Bs2 ds.
Therefore
1
2
Bs3 dBs =
1
4
Bt4
3 t
2 0
Bs2 ds.
(i)
1 2 1
1
P = 2 0 2 ,
4
3 1 0
(ii)
(a)
P133 =
14
7
=
= 0.21875.
64
32
(b)
2
P23
=
2
1
=
= 0.125.
16
8
(c)
14 3
9 3
8 3
14 14 + 9 20 + 8 17
512
8
P13 +
P23 +
P33 =
=
=
.
31 64
31 64
31
31
31
31
(iii)
8 3 5
1
P =
8 6 2 ,
16
5 6 5
2
29 21 14
1
P =
26 18 20 .
64
32 15 17
3
By time n = 300 the effects of the starting point have worn off. The
answer is therefore indistinguishable from the stationary probability 3 in
all three cases.
It is easily observed that the distribution in (c) is stationary, so that
8
3 =
.
31
(i)
The daily change in value of a share is generally on a scale consistent

with the value of the share: this tends to indicate that model II is
preferable.
(ii)
(a)
Model II does appear to fit better than model I; the S dataset does
indeed exhibit large variations when it is at a high level, and
smaller ones when low.
However, the fit does not appear all that good, as Brownian
increments are normally distributed, so are seldom as large as
some of the jumps which appear in this dataset.
Page 3

(b)
If the Brownian model were accurate, the day-to-day increments

sn sn1 should be independently normally distributed with
constant mean and variance:
a test of normality (Anderson-Darling, Kolmogorov-Smirnov, 2
goodness-of-fit) would do fine; a test of independence (based on
sample ACF, or the Durbin-Watson statistic) would also be a good
suggestion.
(iii)
(i)
(a)
A Lvy process is the sum of three independent components: a

deterministic part of the form + t, a Brownian part of the form
Bt and a pure jump part which may be thought of as a compound
Poisson process.
(b)
One problem would be in estimating the distribution of the jump

sizes, particularly with only 250 observations. Even if a family
were assumed for the distribution (e.g. double exponential), there
would be the additional difficulty that small jumps would not be
detectable against the background of the Gaussian noise.
(a)
First the uk need to be transformed so that their distribution is

something suitable for the white noise sequence of a time series,
since at the very least the mean of the sequence needs to be zero.
N (0, 2e ) is the standard choice: one method of achieving this is to
define, for each integer t,
e2t
= e 2 log u2t sin(2u2t +1 )
e2t+1 = e 2log u2t cos(2u2t +1 ),

but there are others, such as the polar method, inverse transform
method or acceptance-rejection sampling.
The values of the et can now be fed into the formula to give the
values of the Xt , whichever model is in use.
(b)
(ii)
Page 4
The ability to re-use a pseudo-random number sequence is

important when comparing the ability of different mechanisms to
control a process which is affected by randomness: in order to
ensure fair comparison of the mechanisms, the must be subjected
to the same degree of random input.
The models do not possess the correct correlation structure.

(iii)
(a)
1 = Corr(Xt ,Xt1) = 1Corr(Xt1 ,Xt1) + 2Corr (Xt2 ,Xt1) = 1 + 2 1 .

Hence 1 = 1 / (1 2)
2 = Corr(Xt ,Xt2) = 1Corr(Xt1 ,Xt2) + 2Corr(Xt2 ,Xt2) = 1 1 + 2 .
(b)
We have 0.7 = 1 = 1 / (1 2)
and 0.5 = 2 = 2 + 12 / (1 2) = 2 + 0.71. Two equations in two
35
1
, 2 =
.
unknowns. Solution: 1 =
51
51
(2 marks for the observation that 1 = 0.7 and 2 = 0 is very close
to giving the right answer, as it gives 2 = 0.49.)
(i)
(t ) = P0,1(t) P0,0(t), or a more general form such as P0,0

(t ) =
P0,0
P0,k(t)k,0
(ii)
Since P0,1(t) = 1 P0,0(t),

(t ) = (1 P0,0(t)) P0,0(t). Any solution method will do,
we have P0,0
d (+)t
[e
P0,0(t)] = e(+)t , solved by P0,0(t) =
+ Ce(+)t, with C
dt
+
being determined by the fact that P0,0(0) = 1.
e.g.
(iii)
E0Ot = E0 t0 Is ds =
=
E0 Is ds =
P0,0(s)ds
t+
(1 e(+)t)
2
+
( + )
(iv)
Since the process must be in state 0 or state 1 at all times, the solution is
t
just t E0 0t =
(1 e(+)t).
+
( + )2
(v)
(a)
Assuming a member who is initially healthy, expected outgoings

(including expenses) by time t and expected income by time t, are
respectively
t +
and
t
(1 e ( + )t )
2
+
( + )
t+
(1 e ( + )t ) .
2
( + )
+
Page 5

In the long run, then, as t , we require = + ( + ) to
break even.
(i)
(b)
The assumptions required are that the rate of becoming ill and
rate of recovery from illness are constant.
(c)
This will certainly not be true of any individual member but, if

membership is large and the age and health profiles of the
members are constant by virtue of a constant influx of new
members, it may be a reasonable approximation.
If Vt is a martingale, then its expectation must be constant and equal to

its initial value euy.
Therefore Eeu( y+t +Bt )c(u )t = euy +(u+u
2 2
/ 2 c ( u ))t
= euy.
Thus we must have c(u) = u + u22 / 2.

(ii)
The optional stopping theorem states that if Mt is a martingale, and T is a

random stopping time, then under some additional technical conditions
(such as MtT being uniformly bounded) we have:
EMT = M0.
It is frequently used to evaluate the expectation of a function of T, such as
the moment generating function (as in this instance).
(iii)
Applying the optional stopping theorem to the martingale Vt we find that

EVTa = euy = Eeua c(u )Ta .
The equation c(u) = v has two roots u+ , u , one being negative and the
other positive (since v is positive).
Now Vt Ta = eu(v )Yt vt and Yt a for 0 t Ta. If u(v) < 0, then
0 < Vt Ta eu(v )a for all t, so that the technical condition is satisfied; the
same cannot be said if u(v) > 0.
Therefore the positive root is unacceptable and f(y, v) = E y e vTa = eu (v )( y a ) .
Comment: For the record, there were two very slight errors in this question, both
appearing as subscripts. In line 4, first formula: T{ } should have read T{a} , and in part
(iii) line 1: T{u} should have read T{a} . This was taken into account by the markers, and
the examiners ensured that no marks were lost by students because of either small error.
Page 6
10
(i)
PAA (s, t) = e
(ii)
P[Rs w] = PAA (s, s + w) = e (( s +w )
xdx
= e (t
s2 ) / 2
E[Rs] =
e sww
/2
s2 ) / 2
= e sww
/2
. Therefore
dw.
Complete the square at the exponent to get

E[Rs] = e s
(iii)
/2
e ( s+w )
/2
dw = e s
/2
e x
/2
dx
1 1 G( s )
g( s )
From (ii) and the given bound

E[ Rs ]
E[ Rs ]
1 1
1
s
1
1
1
3 2.
=
3 3/2
s s
s
The first inequality yields
1
1
=
= 0.00238 . (year)2.
sE[ Rs ]
420
The second inequality can be written as

2E[Rs]
1
+
0,
s s3
so must lie outside the interval:

1
1 4 E[ Rs ]
4E[ Rs ]
24
1 1
1 1
2
3
s
70
s
s
=
=
= [0.00023, 0.00215]
2E[ Rs ]
2sE[ Rs ]
2 6 70
In fact, since clearly
1
we see that must lie in the interval
sE[ Rs ]
[0.00215, 0.00238].
(iv)
Use the inverse transform method, X = F1(U).

x
[ a + bt ] dt ) = exp{a(x 70) b(x2 702)}.
In this case 1 F(x) = exp( 70
Page 7

Therefore bx2 + ax = log(1 F(x)) + 702b + 70a. Replace F(x) by u to
get
x = F1(u) =
a + a 2 + 2b[ log(1 u) + 702 b + 70a ]

b
= r + 702 + 140r + r 2 2b1 log(1 u ),
where r = ab1. If u is an observation of a uniform pseudo-random variate,

then x is an observation from the required distribution.
11
(i)
Consumer prices do tend to exhibit regular seasonal variation, though not

a great deal these days. And, since prices tend to go up rather more than
they come down, it is probably worth including a trend term in any model.
It is certainly possible to test whether the trend term is equal to zero.
(ii)
(a)
Xn+1 xn = (xn xn1) + en+1 + en .
(b)
The parameters are , and 2e . The trend removal process would

have accounted for any parameter.
(iii)
x n (1) = E(Xn+1xn , ..., x1) = xn + (xn xn1) + E(en+1 + enxn , ..., x1). Now
en+1 has mean 0 and is conventionally supposed independent of everything
that happens before n.
On the other hand, en can be deduced from past data,
e.g. en = xn xn1 (xn1 xn2) en1 , which may be iterated back to get
en in terms of the known x and the known e0.
Thus
x n (1) = xn + (xn xn1) + en .
Similarly,
x n (2) = E(Xn+2Fn) = E(Xn+1 + (Xn+1 xn) + en+2 + en+1Fn )
= (1 + ) x n (1) xn .
We see that Xn+1 x n (1) = en+1 , so that the prediction variance is just
Var(en+1) = 2e .
Page 8
(iv)
Since en = xn x n 1 (1) , we have

x n (1) = xn + (xn xn1) + (xn x n 1 (1));
if we set = 0 and = , the equation is identical to the updating
equation for exponential smoothing.
(v)
An ARIMA(p, d, q) model is I(d); in this case, x is I(1).

A stationary (I(0)) model has an equilibrium distribution: the distribution
of the forecast of Xn+k would converge to equilibrium for large k. An I(1)
process is the partial sum of an I(0) process, so would have increasing
variance, even if the mean happened to be stable.
(vi)
Two series {x} and {y} are cointegrated if both are I(1) but there are some
constants a and b such that {ax + by} is stationary.
Two processes are likely to be cointegrated if one drives the other, or if
both are driven by the same underlying process. In the given instance the
suggestion is certainly worth investigating.
Page 9
EXAMINATIONS
10 April 2001 (pm)

1.
2.
3.

separate sheet.

103A2001
{N(t) : t 0} is a Poisson process with rate and {.t : t 0} is the filtration

associated with N.
(i)
Write down the conditional distribution of N(t + s) N(t) given .t , where

s > 0 and use your answer to find E(N(t+s).t).
(ii)
Find a process of the form M(t) = (t)N(t) which is a martingale.
[3]
[2]
[Total 5]
An insurance company wishes to test the assumption that claims of a particular

type arrive according to a Poisson process model. The times of arrival of the next
20 incoming claims of this type are to be recorded, giving a sequence T1 , , T20.
(i)
Give reasons why tests for the goodness of fit should be based on the inter[1]
arrival times Xi = Ti Ti1 rather than on the arrival times Ti .
(ii)
Write down the distribution of the inter-arrival times if the Poisson

process model is correct and state one statistical test which could be
applied to determine whether this distribution is realised in practice.
[2]
(iii)
State the relationship between successive values of the inter-arrival times

if the Poisson process model is correct and state one method which could
be applied to determine whether this relationship holds in practice.
[2]
[Total 5]
(i)
Give a definition of the spectral density of a stationary time series,

expressed in terms of the autocovariance function {k : k Z} of the
process. Use this definition to derive the spectral density of a first-order
moving average process and of a first-order autoregression.
[5]
(ii)
Suppose the inverse of a time series model with spectral density f() is
1
. Using part (i), state
defined to be the model with spectral density
f ( )
the form of the inverse of a first-order moving average and state the way
in which the inverse of an invertible MA(1) differs from the inverse of a
non-invertible MA(1).
[2]
[Total 7]
103 A20012
Let Xn denote an autoregressive high frequency time series modelled by:

Xn+1 = (1 ) Xn + + en ,
where en = 1 with equal probabilities and , , are constant parameters. An
analyst wishes to investigate whether this series may be approximated by some
continuous time diffusion, i.e. Xn Ynh , where Yt satisfies a stochastic differential
equation
dYt = (Yt) dt + dBt
and Bt denotes standard Brownian motion.
(i)
State the expectation and variance of dYt = Yt+h Yt , the increment of the
process Yt over a small interval of size h, conditional on Yt = y.
[2]
(ii)
Calculate the expectation and variance of the increment Xn+1 Xn of the

autoregression, conditional on Xn = y.
[2]
(iii)
Find, by equating the first and second moments of the increments in (i)
and (ii) above, an expression for the drift (y) of the approximating
diffusion in a form which does not involve the time increment h.
[2]
(iv)
State a condition under which the approximating process in (iii) is a

Brownian motion with drift.
(v)
State a condition under which the approximating process in (iii) is an

Ornstein-Uhlenbeck process.
[1]
[Total 8]
(i)
Derive expressions for 1 and 2 , the autocorrelation function of X at lags

1 and 2, in the case that X is a stationary process satisfying the recursion:
[1]
Xt = Xt1 + et + et1 ,
where {et : t = 1, 2, } is a sequence of uncorrelated random variables with
[5]
mean 0, variance 2.
(ii)
A companys monthly sales figures, corrected for trend and seasonal

factors, exhibit sample autocorrelation function at lags 1 and 2 of r1 = 0.5,
r2 = 0.4. Find method of moments estimators of and for the model
in (i).
[3]
[Total 8]
103 A20013
PLEASE TURN OVER
A motor insurance company has 80,000 policy holders, paying an average annual
premium of 400. The company receives claims at a rate of 2000 per month, the
sizes of the claims having mean 1,200, standard deviation = 200.
Let S(t) denote the companys total surplus at time t, with S(0) equal to the
initial reserve, set at 20,000,000.
(i)
Calculate the expectation of the total amount paid out in claims in a given
month and the safety loading employed by the company.
[2]
(ii)
State, with reasons, whether it would be appropriate to use a diffusion

approximation to calculate the probability of ruin, that is the probability
that the process {S(t) : t 0} ever hits 0.
[3]
(iii)
Describe a simulation-based method for estimating the probability of ruin.

Indicate why it would be important to use a method of generating pseudorandom variables which gives rise to reproducible sequences.
[4]
[Total 9]
The evolution of a stock price St is modelled by

St = et +Bt ,
where Bt represents a standard Brownian motion, and are fixed parameters
and the initial value of the stock is S0 = 1.
(i)
Derive an expression for P{St x}.
[2]
(ii)
Derive expressions for the median of St and the expectation of St .
[4]
(iii)
(a)
Determine an expression for the conditional expectation E(StFs),

where s < t and where {Fs : s 0} denotes the filtration associated
with the process S.
(b)
Find conditions on and under which the process {St : t 0} is a

martingale.
(c)
State, with reasons, whether or not the stock would be a good long
term investment in this case.
[5]
[Total 11]
103 A20014
A company assesses the credit-worthiness of various firms every quarter; the

ratings are, in order of decreasing merit, A, B, C and D (default). Historical data
support the view that the credit rating of a typical firm evolves as a Markov
chain with transition matrix
1 2
1 2 2
P=
2
1 2 2
0
0
0
for some parameter

(i)
Draw the transition graph of the chain.
[2]
(ii)
Determine the range of values of for which the matrix P is a valid

transition matrix.
[2]
(iii)
State, with reasons, whether the chain is irreducible and aperiodic.
(iv)
Derive a stationary probability distribution for the chain and establish

whether it is unique.
[4]
(v)
For the value = 0.1, calculate the probability that the companys rating
in the third quarter, X3, is in the default state D:
[2]
(a)
in the case where the companys rating in the first quarter, X1, is
equal to A
(b)
in the case X1 = B
(c)
in the case X1 = C
(d)
in the case X1 = D
103 A20015
[3]
[Total 13]
PLEASE TURN OVER
A continuous-time Markov sickness and death model has four states: H (healthy),
S (sick), T (terminally ill) and D (dead). From a healthy state transitions are
possible to states S and D, each at rate 0.05 per year. A sick person recovers his
health at rate 1.0 per year; other possible transitions are to D and T, each with
rate 0.1 per year. Only one transition is possible from the terminally ill state,
and that is to state D with transition rate 0.4 per year.
(i)
Draw the transition graph for this process.
(ii)
Define P(t) = {pij(t) : i, j H, S, T, D} where pij(t) denotes the probability of

being in state j at time t given that the individual was in state i at time 0.
State the Kolmogorov forward equation satisfied by the matrix P(t),
making sure that you specify the entries of the matrix A which appears.
[3]
(iii)
Calculate the probability of being healthy for at least 10 uninterrupted

years given that you are healthy now.
[1]
(iv)
Let dj denote the probability that a life which is currently in state j will
never suffer a terminal illness. By considering the first transition from
1
state H, show that dH = 12 + 12 dS and deduce similarly that dS = 12
+ 56 dH .
Hence evaluate dH and dS.
(v)
10
[2]
[5]
Write down the expected duration of a terminal illness, starting from the
moment of the first transition into state T. Use the result of (iv) to deduce
the expectation of the future time spent terminally ill by an individual
who is currently healthy.
[4]
[Total 15]
A family agrees an expenditure target, Yn , for year n, in such a way that the
annual increase in the expenditure target is proportional to the increase in the
family income over the previous year. The actual expenditure during the year,
Xn , is assumed to be related to the expenditure target, but incorporating an
element of randomness and a factor accounting for the familys propensity to
overspend. The family income, In , is assumed to grow at a constant annual rate,
before randomness is taken into account.
The head of the household believes that the following three equations form an
appropriate representation of the above information:
Yn = Yn1 + (In1 In2)
Xn = (1 + ) Yn + en(1)
In = (1 + ) In1 + en(2)
where {( en(1) , en(2) ) : n = 1, 2, } is a sequence of zero-mean bivariate Normal
random variables and , and are positive parameters (with < 1).
103 A20016
(i)
Express the first of the equations in terms of the backshift operator, B,

and deduce that a linear relationship exists between Yn and In1.
[3]
(ii)
Show that the process Zn = (Xn , In) is a first-order multivariate

autoregressive process.
[2]
(iii)
State, with reasons, whether {In : n 1} is a stationary time series, and

hence determine whether {Zn : n 1} is I(0), I(1) or neither.
[3]
(iv)
Find an estimator for the parameter by minimising the quantity

tn=2 ( et(2) )2 .
(v)
[3]
The head of the household wishes to perform a simulation to investigate

whether the propensity to overspend will result in negative net savings.
It is assumed that Var( en(1) ) = 12 , Var( en(2) ) = 22 and
Cov( en(1) , en(2) ) = 12 , where 1 < < 1.
(vi)
(a)
Describe a method of simulating an observation of the pair

( en(1) , en(2) ) starting from two uniformly distributed pseudo-random
variables U1 , U2 .
(b)
Describe the role of sensitivity analysis in drawing conclusions

from the simulation.
[6]
An alternative model is proposed, involving the logarithms of the

quantities In , Xn and Yn :
ln Yn = ln Yn1 + ln In1 ln In2
ln Xn = + ln Yn + en(1)
ln In = + ln In1 + en(2)
Discuss whether this model is more suitable than the original model. [2]
[Total 19]
103 A20017
EXAMINATIONS
April 2001
EXAMINERS REPORT
(i)
Given .t we know that N(t + s) N(t) ~ Poisson(s).

Hence E(N(t+s).t) = N(t) e(1)s.
(ii)
Now E((t + s) N(t+s).t) = (t + s) N(t) e(1)s, which needs to be equal to

M(t) = (t) N(t). It follows that (t) = e(1)t.
(i)
The inter-arrival times are much more suitable because they are
independent.
(ii)
They should be exponentially distributed with the same mean.

Kolmogorov-Smirnov, Anderson-Darling or 2 goodness-of-fit test can all
be used.
(iii)
Successive values should be independent.

Regress Xi on Xi1 using ordinary least squares, or fit an AR(1) and test
the 1 parameter for significance (equivalent to Durbin-Watson test).
(i)
Spectral density f() =
1
2
keik, or equivalent.
For MA(1), therefore, we have

f() =
2e
(1 + 2 + 2 cos ).
2
And for AR(1),

f() =
(ii)
2e
1
.
2
2 1 + 2 cos
Clearly from (i) the inverse of the MA(1) is an AR(1), with = and with
a different value of 2e .
The word invertible attached to a MA(1) indicates that the inverse is a
stationary AR(1), whereas a non-invertible MA(1) has as inverse an
AR(1) model which cannot be stationary, such as Xt = 2Xt1 + et.
Page 2
(i)
E (dYt |Yt = y ) = ( y )h + o(h) , Var(dYt |Yt = y ) = h + o(h )
(ii)
E ( X n +1 X n | X n = y ) = y , Var( X n +1 X n | X n = y ) = 2 .
(iii)
(y)h = ( y) and h = 2, so (y) = ( y) / 2.
(iv)
The increments of a brownian motion do not depend on its current value,

i.e. = 0.
(v)
An OU process drifts towards zero, so that = 0.
(i)
Let k denote the autocovariance function of X. Then

Cov (Xt , et) = 0 + 2 + 0 = 2;
Cov (Xt , et1) = 0 + 0 + 2;
2 = 1
1 = 0 + 0 + Cov (Xt1 , et1) = 0 + 2
0 = 1 + Cov (Xt , et) + Cov (Xt , et1) = 20 + (1 + 2 + 2) 2,
implying that
(ii)
0 =
2
(1 + 2 + 2 ) ,
2
1
1 =
( + ) (1 + )
, 2 = 1.
1 + 2 + 2
Estimate of is r2 / r1 = 0.8; estimate of is given by

1
2
(1 + 2 + 2) = ( + ) (1 + ), or 0.3 2 + 0.84 + 0.3 = 0, with solution
= 1.4
0.96 .
In this case we take the positive square root to ensure invertibility.
Page 3
(i)
Expectation is 2.4m. Safety loading is

=
C
(400 /12) 80,000 2,000 1,200
=
= 11.11%
2,000 1,200
where denotes the mean arrival rate of claims and the mean claim
size.
(ii)
Conditions for validity of diffusion approximation are
large, small,
u moderate, where u is the reserve.
In this case
u =
u
is large, is not particularly small and
2 107 11.1 102

1,200
= 1,852, clearly too large. We conclude that the
diffusion approximation is not appropriate.

(iii)
Decide on a quantum of time, which may be a month or may be smaller.

For each time period generate a Poisson variate to indicate the number of
claims received and, conditional on this, a Normal variate with
appropriate mean and variance to represent the total sum claimed.
Subtract this from the total premium income over the period
(deterministic), using the resulting quantity as the increment of the
surplus process. Run the simulation for an extended period of time,
stopping if/when it goes below zero. A large number of simulations should
be performed, with the probability of ruin being estimated using standard
techniques based on the Binomial distribution.
The importance of reproducibility is for sensitivity analysis. The
estimated probability may depend heavily on the values assumed for
mean and standard deviation of the claim size, or on other numerical
parameters. It is necessary to vary the initial assumptions and run the
simulation again, just to ensure that conclusions are not substantially
changed if the parameter values used do not adequately reflect the actual
conditions experienced.
(i)
P{St x} = P{exp (t + Bt) x} = P{t + tN ln(x)} =
ln( x ) t
t
) , where
denotes the standard Normal distribution function.

(ii)
We have to find m so that P{St m} = P { e t +Bt m} = P{t + Bt ln (m)}

= P{Bt ln (m) t} =
1
2
. Since Bt is a symmetric normal variable, its
median is 0 and the last equation can only be satisfied if ln(m) t = 0

and m = et .
The expectation is ESt = Ee(t+B(t)) = Eet e
Page 4
tN
2 t
= et e 2 = e
(+ )t .
2
2
(iii)
By the same token, E(StFs) = S(s) E(e
(ts)+(B(t)B(s))
) = S(s)
(+ )(t s) .
e
2
2
If S is to be a martingale, the conditional expectation must be equal to

S(s).
This will happen if = 12 2 .
From part (ii) we see that for this stock with initial value 1, the median of
the distribution at time t goes to 0 exponentially fast for large t hence, a
very bad investment!
(i)
Transition Graph
1 2
1 2 2
1 2 2
(ii)
All transition probabilities must lie in [0,1].

Now 1 2 2 1 2 1 for 0, so it suffices to ensure
that 1 2 2 0 i.e. 2 1. So the range of possible values of
is [0, 2 1].
(iii)
The chain is not irreducible since D is a trap state.

The chain is aperiodic by inspection.
(iv)
A stationary probability distribution, if it exists, must obey

(1 2) A + B + 2C
A + (1 2 2) B + C
2A + B + (1 2 2) C
2B + C + D
=
=
=
=
A
B
C
D
The last equation implies B = C = 0, and this in turn shows that A = 0.

Hence the stationary probability distribution is = (0, 0, 0, 1)T.
Page 5

It is unique: there is just one recurrent class and it is aperiodic. (Or point
out that there is no other solutions to the equations.)
(v)
With = 0.1, the transition matrix is

0
0.89 0.1 0.01
0.1 0.79 0.1 0.01

0.01 0.1 0.79 0.1
0
0
1
0
Its square is
0.8022 0.169 0.0268 0.002
0.169 0.6441 0.159 0.0279

0.0268 0.159 0.6342 0.18
0
0
1
0
the relevant entries being the last column.
(i)
Transition Graph:
0.05
H
0.1
S
1.0
0.05
0.1
0.4
(ii)
KFE: P (t ) = P(t) A,
0
0.05
0.1 0.05
1.0 1.2 0.1 0.1
A=
.
0
0
0.4 0.4
0
0
0
0
(iii)
Page 6
0.1e0.1x dx = e1.
The probability of staying in state H for 10 years is 10

(iv)
First transition from H must be to S or D, each equally likely. If to D, then

it is certain that no terminal illness will occur; otherwise, the probability
of avoiding a terminal illness is dS .
From S similarly, except that the transition probabilities are to H with
0.1
1
prob. 1.0
= 56 , to D or to T, each with prob. 1.2
= 12
. Once in T it is not
1.2
possible to avoid terminal illness.
Solving the above equations, dS =
dH =
(v)
13
14
1
12
5
6
12 (1 + dS ), implying that dS =
6
7
The Markov property implies that the time spent in state T has
exponential distribution. The rate is 0.4 per year, so the expectation is 2.5
years.
The expected time spent in terminal illness given current health is 2(ever
hit T X0 = H) 2.5 years =
10
2.5
14
(i)
(1 B) Y = B (1 B) I,
with solution Y = BI + const
(ii)
We have the vector equation
years.
X n 0 (1 + ) X n 1 const en(1)
=
+
+ (2) ,
I n 0 1 + I n 1 0 en
which clearly represents a vector AR(1).

(iii)
I is not stationary: the condition for an AR(1) to be stationary is that the

autoregressive parameter is less than 1 in absolute value.
I is not I(1), either, since I = e(2) + BI which, as already stated, is not
stationary.
Z is therefore neither I(0) nor I(1).
(iv)
The equation for the sum of squares is

SS =
( et(2) )2 =
t =2
(I
t =2
(1 + ) It 1 )2 .
Differentiating,
0 = 2
I
t =2
t 1 ( I t
(1 + ) I t 1 ),
Page 7

implying that
=
(v)
(a)
nt=2 It 1 ( I t I t 1 )
tn=2 I t21
First we need to obtain N(0,1) variates Z1 and Z2. The core reading
mentions two methods: either
Z1 =
2 lnU1 sin(2U2),
Z2 =
2 lnU1 cos(2U2)
or
Z1 = V1
2 ln S
,
S
where Vi = 2Ui 1, S =
Z2 = V2
2 ln S
,
S
V12 + V22 and any values of U1 and U2
which give S > 1 are rejected.

Now define E1 = 1Z1 , E2 = 2Z1 +
(b)
(vi)
Page 8
1 2 2Z2 .
Sensitivity analysis applies mostly to the initial assumptions.

Values for the parameters , , , 1 , 2 and must be assumed,
but may not correspond exactly to the actual situation. The head
of household should investigate whether making small changes to
the values used will make large differences to the conclusions.
The revised model still meets the requirements set down at the start of
the problem. It is likely to prove more tractable in that ln I is now a
simple random walk with drift, and is therefore I(1).
EXAMINATIONS

1.
2.
3.

separate sheet.

103S2001 (13.2.01)
A stochastic interest rate model postulates that the base lending rate in month t,
it , follows the model:
it = 5% + 0.9 (it1 5%) + et ,
where et , e2 , is a sequence of independent Normal random variables with mean
0, variance 2.
(i)
(ii)
(a)
Determine an expression for it in terms of et , et1 , et2 , , e1 and i0 .
(b)
Calculate the conditional mean and variance of the base lending

rate in month t given that i0 = 8%.
[4]
Derive an estimator for 2 based on observations i1 , i2 , , in of the base

lending rate in n successive months.
[2]
[Total 6]
Let {B(t) : t 0} be a standard Brownian motion and define

1
B1(t) = tB for t > 0, with B1(0) = 0.
t
(i)
Calculate EB1(t), Var(B1(t)) and Cov{B1(s), B1(t)) for s < t. Deduce that B1
is a standard Brownian motion.
[3]
(ii)
(a)
Show that the two probabilities

P[B(t) < ct for all t 1]
and
P[B(t) < c for all 0 t 1]
are equal to one another where c > 0 is a constant.
(b)
103 S20012
Find an expression for the value of these probabilities by stating

the probability density function of M1 = max0t1 B(t).
[4]
[Total 7]
A stationary stochastic process {Yt : t = 0, 1, } satisfies the relationship

Yt = + 0.8(Yt1 ) 0.4(Yt2 ) + et ,
where {et : t = 0, 1, } is a sequence of independent, zero-mean Normal random
variables with common variance 2.
(i)
Calculate the autocorrelation function, k , and the partial autocorrelation

function, k , of Y for k = 1 and 2.
[5]
(ii)
State, without performing additional calculations, what you would expect

[2]
to find if you were to calculate k and k for larger values of k.
[Total 7]
Consider the simplified model of credit rating of companies in continuous time

described below. There are three ratings which a company can have, A, B and D
(default) and the possible transitions are as follows:
(i)
from A to B with rate 4

from B to A with rate
from B to D with rate 3
Write down the matrix form of Kolmogorovs forward equations as it
applies to this model and verify that the transition matrix P(t) = P(0, t)
given below is a solution:
12 e 2t + 12 e 6t
P(t) = 14 e 2t 14 e 6t
(ii)
e 2t e 6 t
1
2
e 2t + 12 e 6t
0
1 23 e 2t + 12 e 6t
1 43 e 2t 14 e 6t .
[5]
Find the time after which a company starting in state A is more likely to
be in state D than in state A.
[2]
[Total 7]
103 S20013
A motor insurance company assumes that a holder of a provisional drivers

licence will make claims according to a Poisson process with rate X per year,
where X is not fixed but is determined randomly for each driver according to the
density function
f(x) = 2e2x
(x > 0).
(i)
Describe how to simulate an observation X from the density f using a

single pseudo-random variable U assumed uniformly distributed on [0, 1].
[3]
(ii)
Explain how, given the value X generated in (i), you would use a sequence
U1 , U2 , of uniform pseudo-random variables to simulate the number of
claims made in two six-month periods by a provisional driver with mean
claim rate X per year.
[4]
(iii)
Describe a simulation-based method for estimating the conditional

probability that a provisional driver makes 2 or more claims in the second
six months of driving given that no claim was made in the first six
months. [Here the value X is to be assumed unknown.]
[2]
[Total 9]
n
The evolution of a stock price is modelled as a discrete time process Sn = i=
1 Xi ,
where X1 , X2 , are independent, identically distributed random variables with

P{Xi = 1} = p and P{Xi = 1} = q = 1 p. The investment will be liquidated at
either the bankruptcy time T0 (the first time n when the price Sn hits 0) or the
first time TK when the price attains a fixed target K, whichever occurs first.
Let T = min(T0 , TK) denote the liquidation time (the exit time from [0, K]). Let
A = {TK < T0} denote the event that the target is met before bankruptcy and let
pk = P[AS0 = k] denote the probability of this event, given that the initial price is
S0 = k, where k {1, , K 1}.
(i)
By conditioning on the price of the stock at time 1, determine a difference

equation satisfied by pk , k = 1, , K 1.
[2]
(ii)
Assume that p = q =
(iii)
1
2
(a)
Show that Sn is a martingale.
(b)
Derive an expression for pk by applying the optional stopping

theorem to this martingale stopped at T.
[5]
Assume now that p q.

(a)
(b)
103 S20014
Determine a value 1 such that Yn = Sn is a martingale.

[4]
Derive an expression for pk in this case.
[Total 11]
According to an interest rate model which operates in continuous time, the

interest rate r(t) may change only by upward jumps of fixed size ju or by
downward jumps of fixed size jd (where jd < 0), occurring independently according
to Poisson processes Nu(t) (with rate u) and Nd(t) (with rate d).
Let Tu denote the time of the first up jump in the interest rate, Td the time of the
first down jump, T = min(Tu , Td) the time of the first jump. Further, let I be
defined as an indicator taking the value 1 if the first jump is an up jump or 0
otherwise.
(i)
Determine expressions for the probabilities P{Tu > t}, P{Td > t} and P{T > t}
[2]
(ii)
Determine the distribution of I.
[2]
(iii)
Show, by evaluating P{T > t and I = 1}, that I and T are independent
random variables.
[3]
(iv)
Calculate the expectation and variance of the interest rate at time t given
the current rate r(0).
Hint: r(t) = r(0) + ju Nu(t) + jd Nd(t).
(v)
Show that {r(t) : t 0} is a process with stationary, independent

increments.
103 S20015
[2]
[3]
[Total 12]
A company keeps records of quarterly sales figures, {St : t = 1, 2, n} for the most
recent n quarters. It wishes to analyse the records with the aim of predicting the
sales figures in the near future.
The model suggested by the company is:
log St = + t + Q(t) + Xt ,
where {Xt : t = 1, 2, , n} is a stationary time series, Q(t) takes the value 1, 2, 3 or
4 depending on whether the tth quarter is the first, second, third or fourth
quarter of the financial year, and 1 + 2 + 3 + 4 = 0.
(i)
Explain why the company has suggested a linear model for log St rather
than a linear model for St .
[1]
(ii)
Explain the significance of the parameters , and {q : 1 q 4} and give

[4]
a reason for the assumption that 1 + 2 + 3 + 4 = 0.
(iii)
Derive a linear filter Yt = +k2= 2 ak log St+k which has the property that the
filtered series {Yt} does not depend on {q : 1 q 4}.
(iv)
[3]
(a)
Using the filtered series {Yt} obtained in (iii), derive an expression

for Yt in terms of , , Xt , Xt1 ,
(b)
State, with reasons, whether {Yt} is I(0), I(1) or neither.
[4]
[Total 12]
In the Vasicek model, the spot rate of interest is governed by the stochastic
differential equation
drt = a(b rt) dt + dBt
where Bt is a standard Brownian motion and a, b > 0.
(i)
A stochastic process {Ut : t 0} is defined by Ut = eat rt.

(a)
Derive an equation for dUt .
(b)
Solve the equation to find Ut .
(c)
Show that
rt = b + (r0 b) eat +
103 S20016
t
0
e a ( s t ) dBs
[5]
(ii)
State the probability distribution of rt and its limit for large t.
(iii)
Derive, in the case s < t, the conditional expectation E[rt.s], where

{.s : s 0} is the filtration generated by the Brownian motion B.
10
[4]
[3]
[Total 12]
For a given driver, any period j is either accident free (Yj = 0) or gives rise to one
accident (Yj = 1). The probability of having no accident during the next period is
estimated using the drivers past record as follows (all values yj are either 0 or 1):
P[Yn+1 = 0Y1 = y1 , Y2 = y2 , , Yn = yn] = pe ( y1 + y2 +...+ yn ) ,
where 0 < p < 1, 0. The cumulative number of accidents suffered by the
driver over the time from period 1 up to period n is
Xn =
Y .
j
j =1
(i)
Verify that the Markov property holds for the sequence X1 , X2 , , Xn ,

and explain why the sequence Y1 , Y2 , , Yn , does not form a Markov
chain.
[3]
(ii)
Draw the transition graph of the Markov chain X and write down its
transition matrix.
(iii)
Determine, being careful to explain your reasons in each case:

(a)
(b)
(c)
(iv)
(v)
[4]
whether the Markov chain X is time-homogeneous

whether it is irreducible
whether it admits a stationary probability distribution
[4]
Starting from the state Xt = j, calculate the probability of suffering no

further accident for the next n successive periods.
[2]
Suppose you are provided with full claims records for a number of a
companys policy holders.
(a)
Describe a method for estimating the parameters and p.
(b)
Explain how to test the assumption that the probability of an

accident depends only on the cumulative number of accidents, Xn ,
and does not have a direct dependence on n.
[4]
[Total 17]
103 S20017
EXAMINATIONS
September 2001
EXAMINERS REPORT
(i)
(a)
Applying the equation iteratively for t, t - 1, , 1, we obtain

it - 5% = et + 0.9et-1 + 0.92 et-2 + + 0.9t-1 e1 + 0.9t (i0 - 5%).
(b)
The RHS has expectation 0.9t 3%, since each eu has expectation
zero, and has variance s2(1 + 0.92 + 0.94 + + 0.92t-2) = s2
(ii)
(1 - 0.92 t )
.
0.19
We have the expression

et = it - 5% - 0.9(it-1 - 5%),
so that a sensible estimator for s2 would be
1
n
(i
t =2
- 5% - 0.9(it -1 - 5%))2 .
(This is both least squares estimator and maximum likelihood estimator.)
(i)
EB1(t) = tEB(t-1) = 0;
Var B1(t) = t2 Var B(t-1) = t;
Cov (B1(s), B1(t)) = stCov (B(s-1), B(t-1)) = stt-1 = s for s < t.
These are identical to the corresponding quantities for B.
(ii)
(a)
1 c
1
P[B(t) < ct for all t 1] = P B < for all 1
u
u u
= P[B1(u) < c for all u 1]

(b)
We know that the density of M1 is 2f(y) for y > 0, where f is the

standard Normal density.
It follows that the required probability is
P[M1 < c] = 2 c0 f(y)dy = 2F(c) - 1.
Page 2
(i)
g1 = 0.8g0 -0.4g1 + 0 and g2 = 0.8g1 - 0.4g0 + 0

Hence g1 =
4
7
g0 and g2 =
f1 = r1 =
4
7
, f2 =
2
35
g0 , implying that r1 =
r2 - r12
1 - r12
4
7
and r2 =
The ACF will reduce to zero as k increases;

the PACF, however, will be equal to zero for all k > 2.
(i)
The generator is
0
-4 a 4a
A= a
-4a 3a = a
0
0
0
d
dt
= -0.4.
(ii)
The forward equation is
2
35
-4 4 0
1 -4 3
0
0 0
Pt = Pt A.
Calculate the right hand side
e -2at ( -2 + 1) + e -6at ( -2 - 1) e -2at (2 - 4) + e -6 at (2 + 4)
Pt A = a e -2at ( -1 + 12 ) + e -6at (1 + 12 ) e -2at (1 - 2) + e -6 at ( -1 - 2)
0
0
-ae -2at - 3ae -6 at
= -a
e -2at + 32 ae -6 at
2
=
(ii)
d
dt
-ae -2at - 3ae -6at

0
3ae -2at - 3ae -6 at
3a -2a t
e
+ 23 ae -6 at
2
Pt
PAA (t) PAD (t)

1-
-2ae -2at + 6ae -6 at
3e -2 at - 3e -6 at
3 -2 at
e
+ 32 e -6 at
2
t t implies
3 -2at
1
e
= e -2at .
2
2
Hence,
t=
log 2
.
2a
Page 3
(i)
Use the inverse distribution function technique.

We have F(x) = 1 - e-2x, and we require U = F(X) = 1 - e-2X, so that
X = - 12 log(1 - U ) .
(ii)
We need two Poisson pseudo-random variables, Y1 and Y2 , each with

mean m = 12 X .
There are two possible methods for generating Poisson(m):
Keep generating exponential variables with mean 1 until the

cumulative sum exceeds m, then set Y equal to one less than the
number of variables generated.
Draw up a table of the cumulative distribution function Fm of

Poisson (m), use a single uniform U and let Y be the first y such
that Fm(y) > U.
(iii)
Carry out the above procedures a large number of times, independently.

Let N0 be the number of times no claims were made in the first six
months, N0,2+ the number of times no claims were made in the first six
months but two or more in the next six. The required estimate is
N0,2+ / N0.
(i)
Let A = {TK < T0}, U = {S1 - S0 = 1}, D = {S1 - S0 = -1}. Conditioning after
one step, we find:
pk = P[AS0 = k] = pP[AS0 = k U] + qP[AS0 = k D] = p pk+1 + q pk-1 .
(Together with pK = 1, p0 = 0 this could be used to solve pk )
(ii)
(a)
Let Fn = s{X1 , X2 , , Xn}. To show Sn is a martingale we need to

show that the conditional expectation of the increments of Sn is 0.
E[Sn+1 - SnFn] = E[Xn+1Fn] = E[Xn+1] = 0.
(b)
By the optional stopping theorem,

Ek(ST) = S0 = k
On the other hand
Ek(ST) = K pk + 0 (1 - pk) = k,
Page 4

which gives
pk =
(iii)
k
.
K
The exponential process Yn = qSn is a martingale if and only if the

expectation of the factors is 1, i.e. if and only if Eq X1 = (pq + qq-1) = 1. This
equation has two roots q = 1, q = qp .
Applying the optional stopping theorem to the martingale yielded by the
second root gives: qk = qK pk + q0 (1 - pk) and pk =
qk - 1
qK - 1
(i)
P{Tu > t} = e -lut , P{Td > t} = e -ldt , P{T > t} = P{Tu > t} P{Td > t} = e -( lu +ld )t
(ii)
P{I = 1} = P{Td > Tu} = 0 lu e -lut P{Td > t} dt =
(iii)
P{T > t I = 1} = 0 lu e -lut P{Td > t} dt =
lu
lu + l d
lu
lu + l d
e -( lu +ld )t
This is the product of P{T > t} and P{I = 1}.
(iv)
Ert = r0 + (ju lu + jd ld) t and Var rt = ju2 lu t + jd2 ld t
(v)
Take 0 < s < t. Then rt+s - rt = ju (Nu (t + s) - Nu(t)) + jd (Nd (t + s) - Nd (t)).

Now Nu (t + s) - Nu (t) has a Poisson(lu s) distribution, not depending on t
and independent of {Nu (r) : 0 r s}, and Nd (t + s) - Nd (t) has a
Poisson(ld s) distribution, also independent of t and of {Nd (r) : 0 r s}, so
rt+s - rt has the same distribution as rs - r0 and is independent of rs - r0.
(i)
Fluctuations in sales tend to be proportional to sales in the sense that

Var(St+1 -St) St2 , company growth tends to be exponential rather than
linear. (Either of these explanations is sufficient.)
(ii)
m + bt represents a deterministic linear trend in the main sales volume; b

is related to the average annual percentage increase in sales, whereas m is
related to the initial value.
The qq refer to predictable seasonal fluctuations above and below the
average from one quarter to another due to the weather, timing of
Christmas, etc.
Page 5

We assume that Sqq = 0 because any non-zero value could be subsumed in
m; estimation procedures report an indeterminacy if we do not make this
assumption.
(ii)
We need to ensure that each qq has a coefficient of
1
4
(assuming that the
1
8
Xt+2 .
filter coefficients add to 1).

The filter
(iv)
( 18 , 14 , 14 , 14 , 18 ) will do.
We have
Yt = m + bt +
1
8
Xt-2 +
1
4
Xt-1 +
1
4
Xt +
1
4
Xt+1 +
Hence
Yt = Yt - Yt-1 = b +
1
8
(Xt+2 + Xt+1 - Xt-2 - Xt-3) .
Clearly Y is not stationary, if only because it has a trend in the mean.

Y, however, does look stationary, so it is reasonable to claim that Y is
I(1).
dUt = d(eat rt) = aeat rt dt + eat drt
(i)
= eat[art dt + ab dt - art dt + sdBt]

= eat(ab dt + sdBt) .
Hence
= U0 + ab
Ut
t
0
e as ds + s
= r0 + b(eat - 1) + s
t
0
t
0
e as dBs
e as dBs
Thus
= e-at Ut = b + (r0 - b) e-at + s
rt
(ii)
From above, rt is Gaussian

with mean b + (r0 - b) e-at and variance
s2
Page 6
e2a( s -t ) ds = s2
1 - e -2at
.
2a
t
0
e a ( s -t ) dBs .

As t , the distribution of the spot rate is N(b, s2 / 2a).
(iii)
rt
= b + (r0 - b) e-at + se-at
t
0
e au dBu .
Hence, by the martingale property of It integrals

E[rtFs] = b + (r0 - b) e-at + se-at E
= b + (r0 - b) e-at + se-at
s
0
t
0
e au dBuFs
e au dBu
= ea(s-t) rs + b(1 - ea(s-t)).
10
(i)
There is an explicit dependence on the past behaviour of Yj , j n in the

probability distribution of Yn+1 ; hence the Markov property does not hold.
On the other hand
P[Xn+1 = jX1 = i1 , X2 = i2 , , Xn-1 = in-1 , Xn = i]
= P[Yn+1 = j - iY1 = i1 , Y2 = i2 - i1 , , Yn-1 = in-1 - in-2 , Yn = i - in-1]
pe -li
=
-li
1 - pe
if j - i = 0,
if j - i = 1.
This is independent of i1 , i2 , , in-1 .

(ii)
Transition graph
1-p
1 - pe-l
1 - pe-2l
pe-l
pe-2l
Transition matrix:
p 1- p
-l
-l
pe
1 - pe
0
pe -2l
1 - pe -2l
O
O
0
Page 7

(iii)
(a)
Chain is time-homogeneous since transition probabilities

calculated in (i) do not depend on time n.
(b)
It is not irreducible since the number of accidents can never go

down.
(c)
There are no recurrent states, hence there can be no stationary

distribution. Alternatively, a stationary distribution, p, if it exists,
must obey
p0 p = p0
p0 (1 - p) + p1 pe-l = p1
p1 (1 - pe-l) + p2 pe-2l = p2 .
M
Since p < 1 we have p0 = 0, and then p1 = 0 etc. Hence no
stationary probability distribution exists.
(iv)
No new accident:
(pe-jl)n = pn e-njl
(v)
(a)
Maximum likelihood would be very easy in this case: choose l and

p to maximise P{( pe -lxk )1- yk (1 - pe -lxk ) yk }.
(b)
Change the model to:

P[Yn+1 = 0Y1 = y1 , Y2 = y2 , , Yn = yn ] = pe -l( xn ,n ) ,
then test the hypothesis that l(x, n) l(x) for all n.
Page 8
EXAMINATIONS
10 April 2002 (pm)

1.
Enter all the candidate and examination details as requested on the front of your answer
booklet.
2.
You must not start writing your answers in the booklet until instructed to do so by the
supervisor.
3.
4.
Attempt all 10 questions, beginning your answer to each question on a separate sheet.

Hand in BOTH your answer booklet, with any additional sheets firmly attached, and this
question paper.
In addition to this paper you should have available Actuarial Tables and
your own electronic calculator.
103A2002
A stock price {St: t = 1, 2, } is modelled as
St = S0 exp X j ,
j =1
where X1, X2, is a sequence of independent Normal random variables, with

E(Xj) = mj, Var(Xj) = s2j .
In order to perform a martingale analysis it is necessary to find a sequence of
constants a1, a2, such that Yt = a t St is a martingale.
(i)
(ii)
Derive a recurrence equation satisfied by the constants {at: t = 1, 2, }.

[You may use the fact that the moment generating function of N(m, s2) is
M(s) = exp(ms + s2s2).]
[4]
State, with a brief explanation, whether the same answer would be obtained if
[1]
Xj had a non-normal distribution with mean mj, variance s2j .
[Total 5]
An analyst wishes to use a model which is based on Brownian motion, but which does
not become too large and positive for large t. The model proposed is
Xt = Bt e-cBt ,
where Bt is standard Brownian motion, and c is a positive constant.
(i)
Verify that there is an upper bound which X never exceeds.
[2]
(ii)
Use Its Lemma to find dXt.
[3]
(iii)
State, with a brief explanation, whether the suggested model is appropriate for
a process which is asymptotically stationary.
[1]
[Total 6]
103 A20022
(i)
Explain briefly what is meant by a linear trend and by seasonal variation in

respect of a sequence of observed values {x1, x2, , xn} forming a time series.
[2]
(ii)
Describe an operation which can be applied to the data in order to remove

additive seasonal variation of period 2.
[2]
(iii)
Derive an appropriate operation to perform on the process

Xt = exp(a + bt + Zt),
where Z is I(1), in order to obtain a stationary process Yt.
[2]
[Total 6]
Consider the second-order autoregressive process

Yt = -2aYt-1 + a2Yt-2 + Zt
where {Zt} is a zero-mean white noise process with Var(Zt) = s2.
(i)
Determine the range of values of a for which the process Y can be stationary.
[3]
(ii)
Derive the autocovariances 1 and 2 of Y in terms of a and s.
103 A20023
[6]
[Total 9]
PLEASE TURN OVER
The ratio of claims to premium income is calculated annually by a division of a large

insurance company over a period of 30 years. The observed values are plotted against
time in Figure 1a, with the sample autocorrelation function (ACF) plotted in
Figure 1b; the dotted lines indicate the cutoff points for significance at the 5% level.
Figure 1a
Ratio of claims to premium income
1.200
Ratio
1.100
1.000
0.900
0.800
29
27
25
23
21
19
17
15
13
11
0.700
Year, t
Figure 1b
Autocorrelation Function for Ratio
Autocorrelation, r k
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
Lag, k
(i)
Explain which feature of the Figures indicates that differencing is not required
in order to obtain a stationary series.
[1]
(ii)
On the basis of the sample ACF, rk , the companys analyst decides to fit a
first-order autoregressive model to the data. State, with reasons, whether you
consider this to be a reasonable decision and indicate what additional plot you
would require in order to make a firmer recommendation.
[3]
103 A20024
(iii)
The model is fitted and the residuals calculated. The sample ACF of the
residuals is shown in Figure 1c. State what conclusions you would draw from
the plot.
[1]
Figure 1c
Autocorrelation Function for Residuals
Autocorrelation, r k
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
Lag, k
(iv)
The fitted model is

Xt = 0.49541 + 0.5118Xt-1 + et .
(a)
Derive forecasts x30 (1) and x30 (2) for the next two values in the
series.
(b)
Comment on the reliability of the forecasts.
[4]
[Total 9]
A disability benefit scheme is modelled in continuous time by a Markov jump process

with states A (active), T (temporarily disabled), P (permanently disabled) and
D (dead). The transition rates are as follows:
A T: 3l
A P: l
A D: a
T A: 5l
T P: 2l
T D: a
P D: 2a
(i)
Write down the generator matrix of the process.
(ii)
Calculate the probability that, having started in state A, the process has visited
neither T nor P by time t.
[3]
(iii)
Write down the matrix form of the Kolmogorov backward differential

equations and use this to derive a differential equation for pPD(t), the
probability that a scheme member in state P at time 0 will be in state D at
time t.
[2]
(iv)
Solve the equation for pPD(t), the probability that a policyholder, initially
permanently disabled, is dead by t.
[2]
[Total 9]
103 A20025
[2]
PLEASE TURN OVER
(i)
Demonstrate that the random variable

X = -log(U)
has an exponential distribution with mean when U is uniformly distributed
over the range (0, 1).
[2]
(ii)
(a)
Explain how the above can be used to simulate a path of the Poisson
process with intensity l = 4.
(b)
Derive a method for simulating a Poisson random variable with

mean 4.
[4]
(iii)
Describe an alternative method for generating a Poisson random variable with

mean 4 based on the cumulative distribution function.
[2]
(iv)
State, giving reasons, which of the two methods in (ii) and (iii) would be more
efficient if a simulation calls for a large number of Poisson random variables
with mean 4.
[2]
[Total 10]
A company is studying the health records of its longest serving employee in order to
improve its provision for health insurance. Let X(t) = H if the employee is healthy at
time t, X(t) = S otherwise. The available information includes the value of X(t) for all
0 t T.
(i)
Explain the principal stages in the formulation and verification of a stochastic

model for this process.
[3]
(ii)
The company chooses to model X as a two-state time-homogeneous Markov

jump process with transition rates sHS = s, sSH = r.
103 A20026
(a)
State the distribution of a typical holding time in state H and of a

typical holding time in state S assuming the model is valid.
(b)
Write down estimates for the parameters s and r of the model in terms
of quantities which may be derived from the available data.
(c)
Indicate one test which could be used to determine whether the data
support the assumption that the Markov jump process model is
suitable.
[4]
(iii)
103 A20027
Having fitted the model the company discovers that the observed distribution
of holding times in state H does not fit the predictions of the timehomogeneous Markov model; in particular, the mean holding time in state H
between visits to state S appears to be decreasing with t.
(a)
Describe the principal difference between a time-inhomogeneous

model and a time-homogeneous one and indicate whether a timeinhomogeneous model might provide a better fit to the observations.
(b)
Explain why the original model could still be used if the company is
large and has a roughly constant age profile.
[4]
[Total 11]
PLEASE TURN OVER
A motor insurer operates a no claims discount system that has five levels. The
percentage of the basic premium paid by the insured in each level is as follows:
Level
% premium charged
5
4
3
2
1
100
90
80
70
60
Insured motorists move between levels depending on the number of claims in the
previous year. For each policyholder, the number of claims per year follows a
Poisson distribution with mean 0.25.
For those in Levels 2, 3, 4 and 5 at the start of the previous year:
if no claims are made during the previous year, the insured moves down one
level (e.g. from Level 4 to Level 3)
if one claim is made during the previous year, the insured moves up one level
(except those in Level 5 at the start of the previous year, who will remain in
Level 5)
if two claims are made during the previous year, the insured moves up two
levels (except those in Level 5 at the start of the previous year, who will
remain in Level 5 and those in Level 4, who will move to Level 5)
if three or more claims are made during the previous year, the insured moves
to Level 5
For those in Level 1 at the start of the start of the previous year, a no claims discount
protection policy applies whereby they remain in Level 1 if they make one claim. If
they make two claims, they move to Level 2. If they make three or more claims, they
move to Level 5. If they make no claims, they remain in Level 1.
(i)
Determine the transition matrix for the no claims discount system (assuming
that all motorists continue their policies).
[3]
(ii)
A policyholder is in Level 3 for the first year of the policy. Assuming that the
policy is maintained, calculate the probability that at the start of the third year
the policyholder will be (a) in Level 1, (b) in Level 3.
[3]
(iii)
(a)
State conditions under which the probability of being in a particular

state after n years converges as n to some limit which is
independent of the initial state.
(b)
Verify that the conditions are satisfied in this instance.
(c)
Determine the ultimate probability that the insured will be in Level 1.

[8]
103 A20028
(iv)
103 A20029
The insurer suspects that the model used for its calculations may be too
simplistic. Given annual data listing numbers of claims per policy, broken
down by discount level, state which test would be most appropriate to test the
assumption that the distribution of the number of claims per policy per year is
Poisson with mean 0.25.
[1]
[Total 15]
PLEASE TURN OVER
10
(i)
State the Lvy decomposition theorem which describes the constituent parts of
a Lvy process.
[3]
(ii)
Let Mt = exp(-2ab + 2bBt - 2b2t), where Bt is standard Brownian motion and

where a and b are positive constants. Define T as the first time that Mt = 1,
with the definition T = if M never hits the point 1. You may assume that
Mt 0 as t , so that
if M hits 1
MT = 1
0 if M never hits 1
(a)
Show that M is a martingale and that 0 Mt 1 for all 0 t T.
(b)
State the optional stopping theorem and use it to prove that the
probability that the Brownian motion Bt ever hits the line a + bt is
e-2ab.
(iii)
[6]
An insurance company earns premium income at a constant rate c per unit

time. Claims arrive according to a Poisson process with rate l; each claim
may be assumed to be of fixed size k. Let Xt denote the total value of all
claims received up until time t and denote by St the companys surplus at
time t,
St = s0 + ct - Xt,
where s0 is a positive constant. Show that {St: t 0} is a Lvy process and
identify the components of the Lvy decomposition of S.
[2]
(iv)
Calculate the expectation and variance of St.
(v)
The company is interested in the probability of ruin, defined as the probability

that St ever goes below 0. An investigator proposes using a Brownian model
[2]
St* = s0 + mt + sBt,
where Bt is a standard Brownian motion, as an approximation to the surplus
process St.
(a)
Calculate the appropriate values of m and s.
(b)
State the significance of the condition c > kl in this situation.
(c)
Write down the probability that the approximating process S* ever hits
0 assuming that c > kl.
(d)
Outline the principal difference between S and S* and state whether

you consider that the probability obtained in (c) would be an
acceptable approximation to the probability of ruin.
[7]
[Total 20]
103 A200210
REPORT OF THE BOARD OF EXAMINERS ON

THE EXAMINATIONS HELD IN
April 2002
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.
K Forman
Chairman of the Board of Examiners
11 June 2002
(i)
Yk is a martingale, and the values Y1, , Yk are determined by X1, .., Xk, so
E[Yk+1X1, X2, , Xk] = E[Yk+1Y1, Y2, , Yk] = Yk.
Now
k +1
E[Yk+1|X1, X2, .., Xk] = E a k +1.exp X j | X1 , X 2 ,..., X k
j =1
k
= ak+1. E e X k +1 .exp X j | X1, X 2 ,..., X k
j =1
Y
= a k +1.E e X k +1 . k
a
k
We therefore require
a k +1 =
(ii)
ak
= a k exp -m k +1 - s2k +1
2
E e X k +1
The solution depends critically on the moment generating function of a

Normal variable. Non-normal variables have different mgfs, so the answer
obtained would be different.
Answers to Question 1 were disappointing considering how straightforward it
was, suggesting that candidates lacked practice at applications of
martingales.
(i)
d
(be - cb ) = (1 - cb)e-cb = 0 when b = 1/c.
db
Check that this is a maximum:
d2
2
(be - cb ) = (-2c + c 2b)e - cb = -ce -1 when
db
b = 1/c. This is negative, so X is indeed maximised at b = 1/c, giving a
maximum value of e-1/c.
(ii)
It's Lemma has a number of possible forms:

df ( X t ) = f '( X t )dX t + f "( X t )(dX t ) 2 , or
f
f
2 f
df ( X t , t ) = ( X t , t )dt + ( X t , t )dX t + 2 ( X t , t )(dX t ) 2
t
x
x
2
are OK, with (dX t ) = dt.
Page 2
d2
db
(be -cb ) = (-2c + c2b)e-cb. Therefore

dXt = (1 - cBt) e-cBt dBt + (-2c + c2Bt) e-cBt dt.
(iii)
Quite the reverse. If B is large and negative X takes enormously negative

values. X is not in the least stationary.
A common cause of lost marks was failing to check that the turning point was
a maximum.
(i)
A linear trend means that the line of best fit to the data plotted against time
would have a non-zero slope or that there is evidence from the observations
that there is an underlying tendency for the values to increase or decrease with
time at a constant rate.
Seasonal variation is another deterministic component of the mean which
causes E(Xt) to depend on the remainder when t is divided by the period, d; to
spot it from the data, look for recurring patterns in the data or check the
sample ACF.
(ii)
Either use moving averages: set Yt = (Xt + Xt-1), which has had the seasonal
variation smoothed out.
Or use seasonal differencing: set Yt = 2 X t = X t - X t -2 (from the BoxJenkins armoury)
Or use any linear filter Yt = a j X t - j as long as
j
a = a
j
even j
(any such
odd j
filter does answer the question, though it may look very strange)
Or method of seasonal means: estimate a mean for the even-numbered
observations and another for the odd-numbered ones, then subtract these from
the corresponding observations to obtain a set of residuals, which can then be
analysed.
(iii)
Set Y = (log X) = (a + bt + Zt) = b + Zt.

Since it is stated that Z is I(1) it follows that Z is stationary.
Question 3 was generally well answered.
Page 3
(i)
In terms of the backwards shift operator we have

(1 + 2aB - a2B2)Y = Z.
We must find the values of a such that the roots of the polynomial
1 + 2ax - a2x2 lie outside the unit circle.
The roots are
1
1 2 , so we require that
a
2 +1
> 1 and
a
2 -1
> 1 , in
a
other words that a < 2 - 1 .

(ii)
Yt = -2Yt-1 + 2Yt-2 + Zt
Cov[Yt,Yt] = 0 = -21 + 22 + 2
(1)
Cov[Yt,Yt-1] = 1 = -20 + 21
(2)
Cov[Yt,Yt-2] = 2 = -21 + 20
(3)
From (2); 1 = -
2ag 0
(4)
1- a2
Substitute for 1 from (4) into (3)

2 = 2a .
2ag 0
1- a2
5a 2 - a 4
+ 20 = g 0 .
1 - a 2
(5)
substitute for 1 from (4) and 2 from (5) into (1)
0 =
( )
)(1 - 6a + a )
s2 1 - a 2
1+ a2
substitute for 0 from (6) into (4) and (5) to find 1 and 2
1 =
-2as2
(1+ a )(1 - 6a
2
Page 4
+ a4
(5a - a ) .s
=
(1 + a )(1 - 6a + a )
2
and
(6)
(Alternative form for the denominator: 1 - 5a2 - 5a4 + a6.)
Generally well answered, although the exact range of permitted values for a
in (i) caused difficulties.
(i)
If r1, r2, r3, were all close to 1, that would indicate a need for differencing.
That is not the case here.
(ii)
The ACF of an AR(1) is geometrically decreasing. That is approximately the

case here, so AR(1) is not completely unreasonable, but we need a plot of the
sample PACF to check.
(iii)
There is a fairly significant departure from a white noise process. The model
does not appear to fit.
(iv)
(a)
x30 (1) = 0.49541 + 0.5118x30,

x30 (2) = 0.49541 + 0.5118 x30 (1) = 0.74896 + 0.2619x30.
[Numerical values are acceptable here: x30 is in fact equal to 1.01, so
the forecasts would be x30 (1) = 1.012, x30 (2) = 1.014: a margin of
error is acceptable, as x30 must be read off the graph.]
(b)
The forecasts are unreliable:

There are only 30 observations in the series, so the confidence intervals
would be quite wide in any case.
In addition, we have seen that the model is inadequate, casting further
doubt on them.
Many candidates provided interesting and sensible comments on the

data provided, though the problems caused by the small sample size
were not generally recognised.
(i)
The generator is
-a - 4l
l
a
3l
-a - 7l 2l
a
5l
.
0
-2a 2a
0
0
0
0
0
Page 5
(ii)
The required event is that either no jump out of A has taken place by t or that a
jump to D has taken place. This has probability
e-(a+4l)t + (1 - e-(a+4l)t)
(iii)
a
.
a + 4l
The backward equations state P(t ) = AP(t).

We have
d
pPD (t ) = -2apPD(t) + 2a
dt
(iv)
The solution to pPD (t ) = -2apPD(t) + 2a with pPD(0) = 0 is pPD(t)

= 1 - e-2at.
Well done, on the whole. Where there were difficulties they may have been
due to practising more with the time-inhomogeneous version of the equation
rather than the time-homogeneous one.
(i)
Either P(X > x) = P(U < e-4x) = e-4x, so that fX(x) = 4e-4x,
Or probability distribution function F(x) = 1 - e-lx has inverse
1
F-1(y) = - log(1 -y). Hence the inversion method reads:
l
(1)
Generate y from U(0, 1).
1
(2)
Return x = - log(1 - y).
l
1
1
- log y is as good as - log(1 - y) here, since 1 - Y is also U(0, 1).
l
l
(ii)
(a)
Obtain numbers y1, y2, , yn by the above procedure, being outcomes

of random variables Y1, Y2, , Yn independent exponentially
distributed with parameter l = 4.
Put tj = y1 + y2 + + yj and return
xt = 0 if t < t1,
xt = j if tj t < tj+1.
(b)
(iii)
Page 6
Take the value of the Poisson process at time 1.
Calculate F(i) for each i = 0, 1, . Then

(1)
Generate u from U(0, 1).
(2)
Return the smallest i such that u F(i).
(iv)
The method in (ii) requires an average of four uniform r.v.s per Poisson r.v.;
the method in (iii) requires only one. Since the distribution function needs to
be calculated only once, (iii) should be much more efficient.
Very poorly answered, considering that it deals only with simulating standard
exponential and Poisson random variables. Even the inverse distribution
function method was largely misunderstood. There must be the suspicion that
many candidates are not getting as far as Unit 7 in the Core Reading.
(i)
First choose a class of model which might be supposed, for physical reasons,
to provide a reasonable fit to the data. Identify the parameters of the model.
Next estimate the values of the parameters from the data.
See if the observed data match the pattern which would be expected if the
model were accurate and if the parameters had the values given by their
estimates. If not, the model should be revised.
(ii)
(a)
Exponential distribution in each case, with rate s in H, r in S.
(b)
The time spent in state H before the next visit to S has mean s-1.
Therefore a reasonable estimate for s is the reciprocal of the mean
length of each visit: s = (Number of transitions from H to S)/(Total
time spent in state H up until the last transition from H to S), although
it would be equally valid to use the Maximum Likelihood Estimator,
which is (Number of transitions from H to S)/(Total time spent in
state H).
Similarly for r .
(iii)
(c)
Testing whether the successive holding times are independent

exponential variables would be best, and any procedure which does
test this is acceptable. Something like using the c2 goodness-of-fit test
on the even-numbered holding times, then again on the odd-numbered
ones, springs to mind, but there may be other, equally reasonable,
answers.
(a)
For a time-inhomogenous model the transition rates s and r are

functions of t.
It is certainly possible to improve the fit by using a timeinhomogenous model in this instance.
Page 7
(b)
If the age profile is represented by a density function f(a); then the

overall average rate at which a healthy employee falls sick is s =
f (a)s(a)da, roughly constant for all t. The same of course applies
to the overall average rate of recovery. (It is not necessary to write
down the integral to obtain full marks: any explanation which covers
the basic principle will suffice.)
Generally good answers. A number of candidates answered part (i) in the

context of this particular model rather than in general terms and so were only
awarded a reduced number of marks.
(i)
Level at start of this year after:
Level at start
of prev yr
0 claims in
previous yr
1 claims in
previous yr
2 claims in
previous yr
3 or more claims
in previous yr
5
4
3
2
1
4
3
2
1
1
5
5
4
3
1
5
5
5
4
2
5
5
5
5
5
For each policyholder, the number of claims in each year has a Poisson (0.25)
distribution. So
P(0 claims)
= e-0.25
= 0.7788
P(1 claim)
= 0.25 e-0.25
e-0.25
= 0.252 .
2
= 0.1947
P(2 claims)
= 0.0243
P(3 or more claims) = 1 (0.7788 + 0.1947 + 0.0243) = 0.0022
0
0
0.0022
0.9735 0.0243
0
0.1947 0.0243 0.0022
0.7788
Transition matrix P = 0
0.7788
0
0.1947 0.0265
0
0.7788
0
0.2212
0
0
0
0
0.7788 0.2212
(ii)
Page 8
In order to be in level 1 in year 3 the policyholder requires two consecutive

claim-free years. The probability of this is 0.77882 = 0.6065.
A similar argument can be used for the probability of being in level 3 in

year 3, but it may be simpler to calculate the whole vector of probabilities x3.
x1 =
x2 =
x3 =
(0 0 1 0
(0 0 1 0
( 0 0.7788
( 0.6065 0
0)
0) . P =
(0
0.7788 0 0.1947 0.0265 )
0 0.1947 0.0265 ) . P
0.3033 0.0396 0.0506 )
Probability of being in level 3 is 30.33%

(iii)
(a)
The required conditions are that the chain is irreducible and aperiodic.
(b)
Irreducibility: level i can be reached from level j in |j - i| steps;

Aperiodicity: pii > 0 for some i.
(c)
The stationary distribution p will not depend on the starting position.

Require
( p1
p2
p3
p4
p5 ) P =
( p1
p2
p3
p4
p5 )
This gives the following equations:

0.9735 p1 + 0.7788p2 = p1
(1)
0.0243p1 + 0.7788p3 = p2
(2)
0.1947p2 + 0.7788p4 = p3
(3)
0.0243p2 + 0.1947p3 + 0.7788p5 = p4
(4)
and
p1 + p2 + p3 + p4 + p5 = 1
(5)
Solving the simultaneous equations:

from (1) p 2 = 1 - 0.9735 p1 = 0.0340p1 ;
0.7788
substitute for p2 into (2) p3 = 0.0340 - 0.0243 p1 = 0.01244p1 ;
0.7788
substitute for p2 and p3 into (3)
p 4 = 0.01244 - 0.1947 0.0340 p1 = 0.00747p1 ;
0.7788
Page 9
substitute for p2, p3 and p4 into (4)

p5 = 0.00747 - 0.1947 0.01244 - 0.0243 0.0340 p1 = 0.00541p1
0.7788
and substituting for p2, p3 , p4 and p5 into (5) p1 = 0.9440

(and p2 = 0.0321, p3 = 0.0117, p4 = 0.0071 and p5 = 0.0051).
(iv)
A chi-squared goodness-of-fit test is best here.

Very good answers on the whole. Some confusion was caused by the fact that
the states were presented in the reverse of the standard order, but most
candidates coped with this pretty well.
10
(i)
A Lvy process Yt can be decomposed as Yt = y0 + mt + sBt + Nt, where mt is a

deterministic component, sBt a continuous random component (Brownian
motion) and Nt a purely discontinuous component which may be regarded as a
compound Poisson process, independent of the Brownian component.
(ii)
(a)
E(Mt+sFt) = exp(-2ab - 2b2(t + s) + 2bBt) E (e2b( Bt + s - Bt )Ft ). The

independent increment property implies that this conditional
expectation is exp(2b2s).
Therefore E(Mt+sFt) = exp(-2ab - 2b2(t + s) + 2bBt + 2b2s) = Mt.

Ought to check that E(|Mt|) < . Since Mt 0, E(|Mt|) = E(Mt) = e-2ab.
Mt 0 by definition. Since Bt is continuous, it follows that Mt is
continuous. Further, M0 < 1. Therefore Mt cannot exceed 1 without
first passing through 1, which does not happen until time T.
(b)
The optional stopping theorem states that, for any martingale Y and
stopping time T adapted to the same filtration, EYT = Y0 if T is bounded
or Y is bounded or YtT is bounded.
The last of these conditions holds in this case.
We conclude that P(B hits a + bt) = P(M hits 1) = E(MT) = M0 = e-2ab.
(iii)
ct is deterministic, -Xt is a compound Poisson process with constant jump

height -k and the multiplier of the Brownian component is s = 0.
(iv)
Since Xt /k ~ P(l), we have E(-Xt) = -lkt, Var(-Xt) = k2lt. Therefore

E(St) = s0 + (c - kl)t, Var(St) = k2lt.
Page 10
(v)
(a)
In order for the mean and variance to match we require m = c - kl,

s = k l.
(b)
c > kl is the condition for premium income to outstrip outgoings on

average.
(c)
s0 + mt + sBt = 0 if and only if Bt = -
s0 m
- t.
s s
From above, the probability that B ever hits the line a + bt is e-2ab.
Therefore the required approximation to the probability of ruin is
ms
(c - k l) s0
exp -2 20 = exp -2
if c > kl.
s
k 2l
(d)
The difference is that S has a discontinuous component, whereas S*

approximates this with a continuous one.
The difference will be significant if s0 is small and c - kl large, as then
S* will be much less likely to hit 0 than S; in the opposite case the
approximation may be quite reasonable.
Most candidates were not able to attempt every part of this question.
Very few made the connection between part (ii)(b) and part (v)(c).
There seems to be a general lack of confidence when dealing with
martingales or Lvy processes.
Page 11
EXAMINATIONS

1.
booklet.
2.
supervisor.
3.
4.

question paper.
103S2002
Consider the Poisson process Xt defined as a Markov jump process with state space
{0, 1, 2, } and transition probabilities
(lt ) j -i -lt
e
Pij(t) =
( j - i )!
(j i).
(i)
Prove that Xt has independent increments.
[3]
(ii)
Hence prove that Xt - lt is a martingale with respect to the natural filtration Ft

associated with Xt.
[3]
[Total 6]
Yt, t = 1, 2, ., is a time series defined by
Yt - a Yt -1 = Z t + (1 - a ) Zt -1
where Zt, t = 0, 1, , is a sequence of independent zero-mean variables with
common variance s2 and where |a| < 1.
(i)
(ii)
State, giving your reasons, the values of p, d and q for which Y is an

ARIMA(p, d, q) process.
Derive the autocorrelation function rk, k = 0, 1, 2, .
[2]
[6]
[Total 8]
Consider the accident proneness model in which the cumulative number of accidents
Xt suffered by a driver is a Markovian birth process with linear transition rates given
by
(i + 1)b if j = i + 1
si,j =
otherwise
0
(i)
Denoting by ai(t) the probability that a driver who has had no accidents at
time 0 has had at least i accidents by time t, explain why
ai(t + dt) = ai(t) + (ai-1(t) - ai(t))ibdt + o(dt)
as dt 0 and hence derive a differential equation satisfied by ai(t).
(ii)
[3]
Show that
ai(t) = (1 - e-bt)i
and deduce the value of P[X(t) = i|X(0) = 0].
103 S20022
[4]
(iii)
A colleague suggests that a better model would involve transition rates si,i+1(t)
which are dependent on t as well as on i. Comment on this suggestion.
[1]
(iv)
The same colleague proposes the model si,i+1(t) =

proposed model.
(i + 1)b
. Comment on the
t
[1]
[Total 9]
An investor believes that the price of gold increases when the volatility of the equities
market is high and decreases when the volatility is low. The investor therefore wishes
to model the price of gold Xt, as an It process defined by dXt = VtdBt + Vt2 dt , where
Bt is standard Brownian motion and Vt is a measure of market volatility calculated
from the equity price information available at time t.
(i)
Comment briefly on the suitability of this model, mentioning in particular its

behaviour when Vt is large and when Vt is small.
[2]
(ii)
(a)
State Its Lemma.
(b)
Use Its Lemma to find an expression for dMt, where Mt = e-2 X t .
(c)
Deduce that Mt is a martingale.
(iii)
[5]
Jensens Inequality implies that

e-2 E ( X t ) E (e-2 X t ).
Show that E ( X t ) X 0 whatever the value of X0 and comment briefly again
on the suitability of the model.
[2]
[Total 9]
103 S20023
PLEASE TURN OVER
(i)
The classification of stochastic models according to discrete or continuous

time variable, discrete or continuous state space gives rise to a four-way
classification. Give four examples, one of each type, of stochastic models
which may be used to model observed processes.
[2]
(ii)
For each of the following observed processes, identify a type of model which
could be used to model the process, stating which features of the process lead
you to make this choice:
(iii)
(a)
a monthly index of food prices
(b)
an index of prices of shares on the London Stock Exchange, constantly

updated
(c)
the status (active, retired, dead) of a member of a pension scheme
(a)
Discuss briefly the importance of model verification.
(b)
A simple random walk model, whose single parameter q is the

probability of an upwards jump, is to be fitted to a set of observations
{x1, x2, , xn} which have the property that |xi+1 - xi| = 1 for each 1 i
< n. Write down the estimate for the parameter of the model and
describe one test which you would use in the process of model
verification.
[5]
[Total 10]
[3]
The members of a disability insurance scheme are classified as active (A),

temporarily disabled (T), permanently disabled (P) or dead (D). Members are
entitled to benefits when they are in state T or P. For the purpose of analysis the state
of each member is recorded on 1 April each year. It is found that the history of a
typical member evolves in time as a discrete-time Markov chain with transition matrix
A
A 0.75
T 0.5
P 0
D 0
(i)
0.1
0.3
0
0
0.05
0.1
0.8
0
0.1
0.1 .
0.2
Draw the transition graph of the chain and find its stationary probability
distribution.
[3]
(ii)
Calculate the mean duration (in years) of a permanent disability benefit.
[2]
(iii)
Calculate the probability that a member, initially active, is either temporarily

or permanently disabled three years after the start of the scheme.
[3]
(iv)
Calculate the probability that a member, initially active, will never draw any
benefit from the scheme.
[2]
[Total 10]
103 S20024
(i)
Describe the three elements of a linear congruential generator, and set out the
recursive relationship used to generate the pseudo-random number sequence.
[3]
(ii)
The first three numbers produced by a linear congruential generator are 0.954,
0.462 and 0.628. Use these to generate three pseudo-random numbers from
the Pareto distribution with a = 2 and l = 1.
[The density of the Pareto distribution is f(x) =
(iii)
ala
(l + x)a+1
( x > 0).]
[4]
An exponential random variable T with rate parameter l may be simulated

using the formula
1
T = - log U
l
where U is uniformly distributed on [0, 1]. Explain how this can be used to
simulate a path of a Markov jump process {Xt: t 0} which has two states,
H (healthy) and S (sick), with transition rates s from H to S and r from S to H.
Assume X0 = H.
[4]
[Total 11]
103 S20025
PLEASE TURN OVER
An analyst is investigating the extent to which the price of a stock at the beginning
and end of a time interval can be used to provide information about its price during
the interval. The model used is
St = S0 + mt + sBt,
where Bt is a standard Brownian motion. The analyst wishes to investigate the
difference between St and the price St which would be predicted given only S0 and
ST, given by
(T - t ) S0 + tST
St =
, 0 t T.
T
Write down E ( St - S0 ), Var(St - S0), E ( ST - St ), Var(ST -St) and

Cov(St - S0, ST - St), for 0 t T.
[3]
(ii)
Calculate the expectation and variance of St - St .
[4]
(iii)
Find the value of t [0, T] where Var(St - St ) is greatest. Comment on your

answer.
[2]
(iv)
Give two reasons why the model
(i)
St = S0exp(mt + sBt)
might be more suitable than the one used by the analyst and discuss whether
the results of (iii) might have been substantially different if the analyst had
used this model in the first place.
[3]
[Total 12]
103 S20026
Let {xt: 1 t n} be a time series to which an ARMA(1, 1) model is to be fitted.

(i)
Write down the defining equation of an ARMA(1, 1) process, identifying the

parameters.
[2]
(ii)
(a)
Outline the Method of Moments parameter estimation technique.
(b)
State the assumptions underlying the Maximum Likelihood parameter

estimation technique.
(c)
Describe briefly the technique known as backforecasting and explain

why it, or something similar, is necessary in this case.
[5]
(iii)
The model fitted to the data is

xt = 5.67 + 0.61xt-1 + et - 0.23et-1
The most recently observed value in the series is x20 = 8.2, with estimated
residual e20 = -1.38.
103 S20027
(a)
Evaluate estimates x20 (1) and x20 (2) for x21 and x22.
(b)
The simplest form of the method of exponential smoothing used at

time 19 gave a forecast for x20 of 8.37. Assuming the smoothing
parameter is equal to 0.2, find the forecast of x21.
(c)
Give an example of a circumstance in which a form of exponential

smoothing might be expected to outperform Box-Jenkins forecasting in
the prediction of future values of the time series.
[5]
[Total 12]
PLEASE TURN OVER
10
The ticket office at a train station has a single ticket machine that is used by travellers
to purchase tickets.
The machine has a tendency to break down, at which point it must be repaired. The
time until breakdown and the time required to effect repairs both follow the
exponential distribution.
Let P1i(t), i = 0, 1, be the probability that at time t (t > 0) there are i ticket machines
working at the ticket office, given that the ticket machine is working at time t = 0.
(i)
Derive the Kolmogorov forward differential equations for P1i(t) i = 0, 1 in

terms of:
s, where 1/s is the mean time to breakdown for a machine; and

, where 1/ is the mean time to repair a machine
s
- s+r t
1 - e ( ) deduce the value of P11(t).
s+r
[3]
(ii)
Show that P10 ( t ) =
(iii)
The station manager is considering adding a second identical ticket machine,

though there is only one repair team to work on the machines in the event that
both are out of action simultaneously. Assuming that a second machine is
added and operates independently of the first one:
(a)
Write down the generator matrix of the Markov jump process Xt which
counts the number of working ticket machines at time t.
(b)
Derive the Kolmogorov forward differential equations for pi(t), i = 0,

1, 2, the probability that i ticket machines are working.
(c)
Given that, for some t, p0 (t ) =

p1 (t ) =
2rs
2s2
2s2 + 2rs + r2
and p2 (t ) =
r2
2s2 + 2rs + r2
2s2 + 2rs + r2
d
show that
pi (t ) =0 for i = 0, 1, 2.
dt
(d)
103 S20028
[4]
State what conclusions you draw from part (c).
[6]
[Total 13]
EXAMINATIONS
September 2002
EXAMINERS REPORT
Introduction
reasonable.
K G Forman
12 November 2002
Questions involving straightforward applications of Markov Chains and Time Series were
well answered, and the standard of answers to questions about martingales and Brownian
motion is improving from year to year, but candidates appeared to experience unexpected
difficulties in relation to the questions on Markov jump processes.
(i)
Prove this using the Markov property. (Note that this mark can be earned for
use of the property even if the word Markov is not mentioned.)
If s0 < s1, < sn < s < t, then
P[ X t - Xs = j| X s0 = i0, , X sn = in, Xs = i]
= P[ X t = i + j|Xs = i] =
(l(t - s )) j -l(t - s )
e
,
j!
independent of i0, i1, , in.

(ii)
We need to prove that E[ X t | Fs ] = X s .

With s < t we have
E[ X t | Fs ] = E [ X t - X s | Fs ] + E[ X s | Fs ]
= E[ X t - X s ] + X s = l(t - s) + X s .
Thus E[ X t - lt | Fs ] = X s - ls.
The key to part (i) was to use the Markov property; only a few candidates managed to do this
part. Part (ii) was generally well answered.
(i)
Since the equation can be written (1 - aB) Y = (1 + (1 - a)B)Z, the process is

ARMA(1,1), or ARIMA(1,0,1).
(ii)
There are a number of possible ways to calculate the gk, and one way which
sidesteps the gk and calculates the rk directly. The solution presented here is
one of the possible answers; other methods were marked on their merits.
g k = Cov [Yt , Yt -k ] .
Rearrange the time series equation to give Yt = aYt -1 + Zt + (1 - a ) Zt -1
Now Cov[Yt, Zt]
Page 2
= 2
and
Cov[Yt, Zt-1] = .Cov[Yt-1, Zt-1] + Cov[Zt, Zt-1] + (1 ).Cov[Zt-1, Zt-1]

= .2 + 0 + (1 ).2 = 2
Therefore
g 0 = Cov [Yt , Yt ] = a.Cov [Yt , Yt -1 ] + Cov [Yt , Z t ] + (1 - a ) .Cov [Yt , Zt -1 ]
= a.g1 + s2 + (1 - a ) .s 2
g 0 = a.g1 + ( 2 - a ) .s2
(1)
g1 = Cov [Yt , Yt -1 ]
= a.Cov [Yt -1, Yt -1 ] + Cov [ Zt , Yt -1 ] + (1 - a ) .Cov [ Zt -1, Yt -1 ]

= a.g 0 + 0 + (1 - a ) .s2
g1 = a.g 0 + (1 - a ) .s2
(2)
substitute for 0 from (1) into (2)
g1 = a. a.g1 + ( 2 - a ) s 2 + (1 - a ) .s 2
( 2 - a ) .a + (1 - a ) .s2 1 + a - a 2 2
g1 =
=
.s
1 - a 2
1- a2
substitute for 1 back into (1)
(1 + a - a ) .s
= a.
2
g0
1- a2
2 - a2 2
+ ( 2 - a ) .s 2 =
.s
1 - a 2
For k 2,
g k = Cov [Yt , Yt - k ]
= a.Cov [Yt -1, Yt - k ] + Cov [ Zt , Yt - k ] + (1 - a ) .Cov [ Zt -1, Yt - k ]

= a.g k -1 + 0 + 0
g k = a k -1.g1
Page 3
The autocorrelation function is: rk =
gk
. Therefore
g0
r0 = 1
r1 =
1+ a - a2
2 - a2
r k = a k -1r1
k2
Part (i) was generally well answered, using a variety of different approaches. Candidates
generally made good attempts at part (ii), the main problems occurring being a failure to
correctly specify 0 and algebraic errors in solving the simultaneous equations.
(i)
If the driver is to have had at least i accidents by time t + dt, either there must
have been i accidents by time t or there must have been exactly i - 1 by time t
and another between t and t + dt.
P (exactly i - 1) = P(at least i - 1) - P(at least i ).
Therefore
dai
= ib(ai-1 - ai).
dt
(ii)
Verification:
d
{(1 - e-bt)i}= ibe-bt(1 - e-bt)i-1 .
dt
ai-1 - ai = (1 - e-bt)i-1(1 - [1 - e-bt]).

We should also verify that ai(0) is correct: the value should be 0 for i > 0,
which it is.
Then, defining Ti as the time the process first hits i, we have
P0,i(t) = P {}
Ti t} - P{Ti +1 t} = (1 - e-bt)i - (1 - e-bt)i+1 = e-bt(1 - e-bt)i
(iii)
Probably a good suggestion. A driver who has had 2 accidents in 10 years is

less likely to have another than a driver who has had 2 accidents in a month.
(iv)
The proposed model does address the issue raised in (iii) but leads to a very
high accident rate when t is close to 0, so is unsuitable.
Many candidates attempted part (i), but were unable to give a sufficiently clear description to
convincingly display their understanding and so did not earn full marks. In some cases,
Page 4
candidates did not correctly interpret the definition of ai(t). In part (ii) many candidates tried
to solve the differential equation, where they could instead have simply shown the solution
given to be valid. Candidates generally came up with sensible suggestions for part (iii), but
only a few candidates were able to make suitable comments on part (iv).
(i)
When volatility is high, dXt is strongly upwards; when volatility is low, dXt is
close to zero.
The first of these fits the assumptions, whereas for the second something more
negative would be preferable. (It looks as though the process X has more
opportunity to increase than to decrease.)
(ii)
(a) It (time-independent case): if dXt = Ytdt + ZtdBt then

df
d2 f
(Yt dt + Z t dBt ) + 2 Zt2 dt. or simply
dx
dx
2
f '( X t )dX t + f ''( X t )(dX t ) , with an explanation of what is meant by (dXt)2.
df(Xt) =
Equally acceptable is the time-dependent version:

df(Xt,t) =
f
f
2 f
dt + (Yt dt + Zt dBt ) + 2 Zt2 dt.
t
x
x
(b) Here
df(Xt) = f ( X t )dX t + f ( X t )(dX t ) 2
= f ( X t )[Yt dBt + Yt2 dt ] + f ( X t )Yt 2 dt
= f ( X t )Yt dBt + Yt2 [ f ( X t ) + f ( X t )]dt.
In this instance, df(Xt) = -2e-2 X t Vt dBt .
(c) This is a martingale by the disappearance of the dt term, as E(dBt | Ft) = 0.
Alternatively,
T
e-2 X T = e-2 X 0 - 2 e-2 X t Vt dBt ,

0
which is a martingale.
(iii)
e-2 X 0 = E ( M 0 ) = E ( M t ) = E (e -2 X t ) e -2 E ( X t ) . It follows that E ( X t ) X 0 ,

whatever the initial state.
Page 5
This confirms the initial suggestion that downward movements are too
unlikely in comparison to upward ones.
The question generally showed good attempts, with a variety of different but correct versions
of It's Lemma being given. The only real problem evident here was in part (iv), where many
candidates failed to use the given inequality to prove the point.
(i)
Discrete state space, discrete time: Markov chain, simple random walk,
anything like that;
Discrete state space, continuous time: Poisson process, Markov jump process;
Continuous state space, discrete time: time series, general random walk;
Continuous state space, continuous time: It process, Brownian motion, etc.
(ii)
(a) by definition, is monthly. Therefore a continuous time variable is not

appropriate. Something like a time series would do, containing an element of
Autoregression.
(b) functions in continuous time; Brownian motion, Geometric BM or any
kind of diffusion or It process would be a suitable candidate.
(c) has a discrete state space. It would be possible to review the members
status only once a year, say in which case a Markov chain would fit, but in the
absence of a remark about frequency of membership status review a
continuous-time model would seem more appropriate; a Markov jump process
is what we might look for.
(iii)
(a)
Once a model has been decided upon, parameters may be estimated by

standard methods. But it is necessary to check that the model, with
parameters given by their estimated values, has sample paths which
resemble the data actually observed, otherwise incorrect inferences can
be drawn.
(b)
Let yi = xi - xi-1 = xi and let nu be the number of time yi = 1. The

parameter, q (or p) is the probability of an up-jump, estimated as
q = nu / (n -1).
We need to check that the yi form a sequence of i.i.d. variables. A test

based on chi-squared is not appropriate except as indicated below.
The test to apply is one which determines whether there is significant
clustering of the +1s and the -1s, as opposed to their being randomly
scattered through the sample: the runs test on the yi, or equivalently the
turning points test on the xi, would be fine; a test based on the sample
autocorrelation function or a contingency table (yi = -1 or +1 as the
Page 6
row labels, yi+1 = -1 or +1 as the column labels) would also be

acceptable.
In part (i) many candidates gave specific examples of observable processes, rather than
describing the stochastic models themselves as requested in the question; some credit was
however given for these cases. Part (ii) and (iii) (a) generally had good answers. In part
(iii) (b) many candidates incorrectly suggested that a chi-squared test should be used,
(i)
(For full credit the probabilities should all be included on the diagram.)
Stationary distribution pA = pT = pP = 0, pD = 1. Derivation not required: the
answer is obvious to anyone who understands.
(ii)
The duration of a permanent disability benefit is a geometric r.v., T.

Since P[ X n +1 = P | X n = P] = 0.8, the parameter is 0.2:
P[T = n] = (0.8)n-10.2. Accordingly its mean is
1
= 5.
1 - 0.8
(iii)
The required probability is obtained from (1 0 0 0)P3.

(1 0 0 0)P = (.75 .1 .05 .1); (1 0 0 0)P2 = (.6125 .105 .0875 .195);
(1 0 0 0)P3 = (.511875 .09275 .111125 .28425).
The solution is .09275 + .111125 = .203875.
(The summation does not have to be performed to earn the mark: giving the
two probabilities separately is a reasonable interpretation of the question.)
(iv)
The probability of never visiting T or P starting from A is

0.1 + 0.75 0.1 + (0.75)2 0.1 +
Page 7
= 0.1 (0.75)n =
n=0
0.1
= 0.4.
1 - 0.75
Most candidates correctly gave the transition graph in part (i), although there were a few
cases where the transition probabilities were not included in the solution. Part (iii) was well
answered. Solutions to (ii) and (iv) were mixed, with perhaps less than a third of students
immediately recognising the required approach and thereby gaining full marks.
(i)
The three elements are:
the multiplier usually denoted a
the increment usually denoted c; the increment is often set to zero

(without any loss in the quality of the pseudo-random sequence) to speed
up the generation process
the modulus usually denoted m where m > a and m > c; the

generator will produce a series of pseudo-random numbers with period no
more than m, so the modulus is usually set to as high a number as possible.
The recursive relationship is: Xn+1 = (aXn + c) (mod m) for n = 0, 1, 2, ,

N - 1. We then set xk = Xk/m for each k = 1, 2, , N.
(ii)
Use the inverse transform method:
f(x) =
ala
(l + x)a+1
x=
(iii)
1
, so that F(x) = 1 -
3
(1 + x)
1+ x
1
- 1. , so
1 - F ( x)
if F(x) = 0.954, then x = 3.6625

if F(x) = 0.462, then x = 0.3634
if F(x) = 0.628, then x = 0.6396
The structure of a Markov jump process implies that the time until the next
jump has exponential distribution, with rate s if the current state is H, r if S.
The even-numbered inter-jump times will have one distribution, the oddnumbered ones a different one.
Obtain numbers y0, y2, y4, , y2n by the above procedure, being simulated
outcomes of independent random variables Y0, Y2, Y4, , Y2n exponentially
distributed with parameter s. Similarly, obtain y1, y3, , y2n-1 using a
parameter r instead of s.
Page 8
Put tj = y0 + y1 + + yj. Find j such that tj-1 t < tj and return

H if j is even
xt =
S if j is odd
Candidates made reasonable attempts at part (i), although in many cases insufficient detail
was provided to score full marks. Part (ii) was generally well answered, the only problem
here being simple algebraic errors. Part (iii) was not answered well, with candidates on the
whole failing to recognise that the transitions from healthy to sick and from sick to healthy
should be modelled separately and then combined to provide the required process.
(i)
E ( St - S0 ) = mt , Var( St - S0 ) =s2t.
E ( ST - St ) = m (T - t ), Var( ST - St ) = s2(T - t).
Cov(St - S0, ST - St) = 0, because of the independent increment property of
Brownian motion.
(ii)
The expectation is mt - (t/T)mT = 0.

Var(St - St )
T - t
t
= Var
( St - S0 ) - ( ST - St )
T
T
t (T - t )s 2
T -t 2 t 2
=
.
s t + T s (T - t ) =
T
T

(iii)
The maximum of t(T - t) is attained at t = T.

This is not surprising. The graph of St - St against t is tied down at the ends,
as the function is constrained to be equal to zero. The greatest scope for
variation is bound to be in the middle.
(iv)
Two possible reasons might be that a Brownian motion can become negative,
which a stock price cannot, and that fluctuations in the value of a stock price
are usually proportional to the price. Other reasons could also apply.
Under the revised model ln(St) has the same structure as St in the original
model, so ln(St) will have its greatest variability at T. The result will not
differ greatly from the result above.
Candidates generally gave good answers for part (i). However, answers to part (ii) and (iii)
were disappointing, particularly for part (iii) where the answer can be derived very simply
from general reasoning. Part (iv) in general showed better answers, although many
candidates failed to get full marks because they did not discuss how their response to (iii)
would differ under the new model.
Page 9
(i)
Xt = m + a(Xt-1 - m) + et + bet-1 is the equation.

The parameters are a (the autoregressive parameter), b (the moving average
parameter), m (the mean level) and s (the innovation standard deviation).
(ii)
(a)
Calculate theoretical ACF r1, r2 of ARMA(1, 1) in terms of a and b.

Find sample ACF r1, r2 from the data. The required estimates are the
values of a and b which ensure that r1 = r1 and r2 = r2.
The value of s2 is estimated using for example the calculated value of
g0 and the sample variance.
(b)
The assumptions are that the en are independent and Normally

distributed.
(c)
MLE in this case tries to minimise et2 . Now et can be expressed as

a function of xt, xt-1 and et-1. et-1 is unknown, but can be expressed as
a function of xt-1, xt-2 and et-2, etc. We need to estimate a suitable
value of e0.
This is done iteratively: assume that e0 = 0 and estimate parameters on
that basis; then use forecasting techniques on the time-reversed process
{xn, xn-1, , x1} to gain a more accurate estimate of e0; repeat this
process until everything converges.
(iii)
(a)
x20 (1) = 5.67 + 0.61(8.2) + 0 - 0.23(-1.38) = 10.99.

x20 (2) = 5.67 + 0.61(10.99) + 0 - 0 = 12.37.
(b)
For exponential smoothing the equation is

x20 (1) = x19 (1) + a ( x20 - x19 (1)) = 8.37 + 0.2(-0.17) = 8.34.
(c)
A variety of exponential smoothing might be better if the mean

changes by some means other than a linear trend, or if there is
multiplicative seasonal variation.
In part (i) a number of candidates omitted from the list of parameters. Part (ii) (a) showed
some good answers, however in most cases insufficient details was provided to earn full
marks. Part (ii) (c) was very poorly answered, with very few candidates showing that they
understood the concept of backforecasting. In part (iii) most candidates understood what
was generally required, though in some cases there were errors in applying the formulae.
10
(i)
P10 ( t + h ) = (1 - rh ) .P10 ( t ) + s.h.P11 ( t )
Page 10
d
P10 ( t ) = - r.P10 ( t ) + s.P11 ( t )
dt
(1)
and
P11 ( t + h ) = r.h.P10 ( t ) + (1 - s.h ) .P11 ( t )
d
P11 ( t ) = r.P10 ( t ) - s.P11 ( t )
(2)
dt
[(2) also follows from the fact that P10 ( t ) + P11 ( t ) = 1 .]
(ii)
P10 ( t ) + P11 ( t ) = 1
so from (1)
d
P10 ( t ) + ( s + r ) .P10 ( t ) = s
dt
d ( s+r )t
s+r t
e
P10 ( t ) = s.e( ) + C
dt
P10 ( t ) =
s
s+r t
- s+r t
.e( ) + C.e ( )
s+r
P10 ( t ) =
s
- s+r t
. 1- e ( )
s+r
since P10(0) = 0.
[Alternatively, instead of solving the DE, just verify that the function
proposed as the solution does indeed satisfy the DE and also check that the
value is correct at t = 0.]
Therefore
P11 (t ) = 1 (iii)
(a)
- s+r t
s
r + se ( )
- s+r t
. 1- e ( ) =
.
s+r
s+r
The generator matrix is now

r
0
-r
s - (s + r) r
0
-2s
2s
(b)
Hence the Forward Equations are

d
p0 ( t ) = - r. p0 ( t ) + s. p1 ( t )
dt
d
p1 ( t ) = r. p0 - ( s + r ) p1 ( t ) + 2.s. p2 (t )
dt
Page 11
d
p2 ( t ) = r. p1 ( t ) - 2s. p2 ( t )
dt
(c)
Simply substituting in the suggested values gives the required result.
(d)
The implication is that the given distribution is stationary. By the

standard properties of Markov processes, it follows that it is the
equilibrium distribution, so that the long-term probabilities of being in
each of the three states are known.
Answers to this question were on the whole disappointing.

Problems in part (i) included transposing the parameters. Marks were available for deriving
p11 from p10 (using the information given in the question) even where p10 could not be
correctly identified.
Only a small proportion of candidates managed to provide a solution for the differential
equation in part (ii).
Better answers were given for the generator matrix in part (iii).
Candidates that gave a generator matrix by and large were able to score marks by applying
their matrix to produce the forward equations. Very few candidates provided any
conclusions under part (iv).
Page 12
EXAMINATIONS
8 April 2003 (pm)

1.
booklet.
2.
supervisor.
3.
4.

question paper.
103A2003
A science student is observing a mouse in the following simple maze:
4
2
1
5
3
6
During each period under observation, the mouse moves to any adjacent, accessible
compartment with equal probability. Successive moves have the Markov property.
Let Xt be the compartment number that the mouse is in at time t.
(i)
Draw a transition diagram, including on your diagram the probabilities of the

possible transitions.
[2]
(ii)
Evaluate the probability that a mouse which starts in compartment 1 at time 0

is also in compartment 1 (a) at time 1, (b) at time 2, and (c) at time 3.
[2]
(iii)
State with reasons whether:

(a)
(b)
the Markov chain X possesses a stationary distribution

P(Xt = j | X0 = i) converges to some limit pj as t
[2]
[Total 6]
An index of salaries, St, and an index of prices, Pt, are modelled as being related to
each other in the following way:
ln St = qS + a1 ln St -1 + a 2 ln Pt -1 + eS ,t
ln Pt = q P + b1 ln St -1 + b2 ln Pt -1 + eP ,t
where eS,t and eP,t are two independent zero-mean white noise processes, with
variances s12 and s22 respectively, and qS and qP are constants.
(i)
Explain why the above model is written in terms of ln P and ln S instead of

just P and S.
[2]
(ii)
Comment on whether it is reasonable that ln St should be affected by

ln Pt-1 and that ln Pt should be affected by ln St-1.
[1]
(iii)
Use matrix notation to express (ln St, ln Pt)T as a Vector Autoregression

and identify the order p of the VAR(p) process.
[2]
(iv)
Suppose the parameters of the model have been estimated. Describe the use of
sensitivity analysis in determining the validity of the model.
[2]
[Total 7]
103 A20032
(i)
Let Yt denote Brownian motion with drift m and variance rate s2, starting at 0.
Write down the expectation E(Yt) and variance Var(Yt) of this process.
[1]
(ii)
Let Xt denote a biased random walk which moves up or down by D at time

intervals of size h, i.e.
Xt =
t
h
Z
i =1 i
where Z1, Z2, are independent random variables, each with distribution
P[Zi = D] = p, P[Zi = -D] = 1 - p. Calculate the expectation and variance of
Xnh and derive an expression for the expectation and variance of Xt.
[3]
[Notation: x denotes the integer part of x.]
(iii)
(i)
(a)
Use the approximation x x to derive conditions on p and D such

that the expectation and variance of the process Xt approximate for
large t those of a Brownian motion with drift m and variance rate s2.
(b)
Comment on the values of h for which such an approximation might be

appropriate.
[4]
[Total 8]
Given a pseudo-random number U uniformly distributed over [0,1], obtain an

expression in terms of U and q for a non-negative pseudo-random variable X
which has density function
f ( x) = qe-qx
(ii)
[2]
A sequence of simulated observations is required from the density function

g ( x) = k (q)
e -qx
, x>0
1+ x
where q is a non-negative parameter and k(q) is a constant of integration not

involving x.
103 A20033
(a)
Describe a procedure that applies the Acceptance-Rejection method to

obtain the required observations.
(b)
Derive an expression involving q and k(q) for the expected number of

pseudo-random variables required to generate a single observation
from the density g using this method.
[6]
[Total 8]
PLEASE TURN OVER
(i)
Let W(t) be defined by W(t) = -4B(kt), where B(t) is a standard Brownian

motion.
(a)
Calculate the value of k which gives W(t) the same expectation and
covariance function as B(t).
(b)
Prove that, for this value of k, W is a standard Brownian motion.

[4]
(ii)
Prove that exp[2B(t)-2t] is a martingale.
[5]
[Total 9]
A No-Claims Discount system operated by a motor insurer has four levels:

Level 1:
Level 2:
Level 3:
Level 4:
0% discount
25% discount
40% discount
50% discount
The rules for moving between these levels are as follows:

Following a claim-free year, move to the next higher level, or remain at
level 4.
Following a year with one or more claims:
move back one level, or remain at level 1, if, in the year before the
most recent year, there were no claims;
move back two levels, or move to level 1 or remain at level 1 if, in the
year before the most recent year, there was one or more claims.
For a given policyholder the probability of no claims in a year is 0.8.
(i)
(ii)
Let X(t) denote the state, either 1, 2, 3 or 4, of the policyholder in year t.

Explain why { X (t )}t=1 is not a Markov chain.
[2]
(a)
By increasing the number of states, define a new stochastic process

{Y (t )}t=1 which is Markov and is such that Y(t) indicates the discount
level for the policyholder in year t.
(b)
Write down the transition matrix for the Markov chain {Y (t )}t=1.
(c)
Calculate the long-run probability that the motorist is in discount

level 3.
[8]
[Total 10]
103 A20034
The medical insurance division of a large insurance company models each

policyholders state of health as a three-state Markov jump process, the states being
H (healthy), S (sick) and D (dead). The instantaneous transition rates between the
states are sHS = s, sSH = r, sHD = m, sSD = n.
(i)
Write down the generator matrix of the Markov jump process.
[2]
(ii)
The policyholder pays contributions at rate C when in state H and receives

benefits at rate B when in state S. No death benefit is payable. The company
uses the model to set the ratio of contributions to benefits. Without doing any
calculations, explain in general terms how this can be done.
[2]
(iii)
A trainee believes that the model is too simplistic. For each of the trainees
suggestions below, comment on whether following the suggestion would be
likely to improve the models predictive power:
(a)
The transition rates should depend on the age of the policyholder.
(b)
The transition rates should vary according to the time of year.
(c)
sSH and sSD should also depend on the duration of the sickness to date.
[3]
(iv)
Outline the principal difficulty in fitting a model with parameters dependent

on all the factors in part (iii).
[1]
(v)
Assume that several years of quarterly claims data are available. Describe a
test to determine whether the model with annually time-varying transition
rates, as in (iii)(b), is a better fit to the data than the model with constant
transition rates.
[2]
[Total 10]
103 A20035
PLEASE TURN OVER
A branch of a bank has three cash dispensers. If at time t a cash dispenser is working,
the probability that it will break down in (t, t + dt) is independent of the state of the
other cash dispensers and is equal to
dt + o(dt).
When a cash dispenser breaks down, repair work begins immediately. The time taken
to repair a broken machine is exponentially distributed with mean 1/. There are
enough repair teams to repair all three cash dispensers at the same time, if necessary.
Define Xt as the number of machines not working at time t.
(i)
Write down the state space for Xt.
(ii)
(a)
If a cash dispenser is working at time 0, prove that the time until its
first breakdown is exponentially distributed with mean 1/.
(b)
If all three cash dispensers are working at time 0, derive the

distribution of the time until the first breakdown.
(iii)
[1]
[5]
Define Pm(t) as the probability that m machines are not working at time t.
Show that the forward equations for the process X imply that:
P0 (t ) = - 3aP0 (t ) + bP1 (t )
P3(t ) = - 3bP3 (t ) + aP2 (t )
and for m = 1, 2,
Pm (t ) = - ((3 - m)a + mb) Pm (t ) + (4 - m)aPm-1 (t ) + (m + 1)bPm+1 (t ) .
(iv)
[3]
At time 0 all three cash dispensers are working. Show that the forward
equations have a solution
3
Pm (t ) = q(t )m [1 - q(t )]3-m , m = 0,1, 2,3
m
where q(t) satisfies a differential equation which you should identify.

[5]
[Total 14]
103 A20036
In a simple discrete-time model for the price of a share, the change in price at time t,
Xt, is assumed to be independent of anything that has happened before time t and to
have distribution:
with probability p
1
,
Xt = -1 with probability q
0 with probability r = 1 - p - q
where p, q, r > 0.
Let S0 = m (where m is a positive integer) be the original price of the share,
Sn = S0 +
n
t =1
X t be the price after n time units and define
Yn = ( q / p ) S n
(i)
Show that {Yn : n 1} is a martingale and that for any positive integer n,
E(Yn) = (q/p)m
(ii)
[4]
Let T be the time until the share price reaches either 0 or N for the first time,
where N is an integer greater than m.
(a)
Show that Yn C for n T, for some constant C.
(b)
Write down an expression for E(YT).

[2]
(iii)
Assuming that p q, calculate the probability that, starting from m, the share
price reaches 0 before it reaches N, i.e. P(ST = 0 S0 = m).
[3]
(iv)
In the case p = q you may assume that P(ST = 0 S0 = m) = (N-m)/N. Now

define Zn = Sn2 - 2np.
(a)
Prove that {Zn : n 1} is a martingale.
(b)
Show that the expected value of the time until absorption, T, is given
by
E(TS0 = m) =
m ( N - m)
.
2p
[5]
[Total 14]
103 A20037
PLEASE TURN OVER
10
The price of a stock through time, {Xt : t 0} is thought to be modelled by the

following relationship:
Xt = 1.7Xt-1 - 0.4Xt-2 - 0.3Xt-3 + et - 0.7et-1 + 0.12et-2

For this model,
(i)
Write the equation in terms of the backward shift operator B in the form
f(B)(1 - B)d Xt = q(B)et,
where f(B) and q(B) are polynomials in B.
[3]
(ii)
Identify the values of p, d and q for which X is an ARIMA(p, d, q) process. [1]
(iii)
Explain whether the process {Xt : t 0} is stationary.
(iv)
For the value of d from (ii), put Wt = (1 - B)dXt. Explain why the model can
be written in the equivalent form
Wt =
yi et -i
and calculate yi for i = 0, 1, 2.

(v)
[1]
[3]
Another representation of the model is

p(B)Wt = et
where p(B) = 1 -
(vi)
i=1 pi Bi .
Calculate pi for i = 1, 2.
[1]
Define two vector-valued stochastic processes Y and Z as

Yt = (Xt, Xt-1, Xt-2)T, Zt = (Xt, Xt-1, Xt-2, Xt-3, Xt-4)T.
Explain which, if any, of the processes {Xt : t 0}, {Yt : t 0} and {Zt : t 0}
has the Markov property.
[2]
(vii)
103 A20038
Given a set of observations (x1, x2, , xn) from an ARIMA(1, 1, 1) process

with unknown parameter values, outline the main steps that need to be taken
so that one can obtain forecasts for future values of the process.
[3]
[Total 14]
REPORT OF THE BOARD OF EXAMINERS

April 2003
EXAMINERS REPORT
Introduction
reasonable.
J Curtis
3 June 2003
(i)
Transition diagram:
1
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
0.5
4
0.5
0.5
(ii)
(a) and (c) are both zero, as it is not possible to return to 1 in an odd number of
steps.
1 1 1
(2)
= p12 p21 + p13 p31 = + = .
For (b): p1,1
4 4 2
(iii)
(a)
Yes every chain with finite state space has a stationary distribution.
(b)
No. The chain is periodic, so the probabilities do not converge.
It was possible to read the question as implying the possibility that the mouse could
stay where it was: candidates who did this were still able to obtain full marks for the
question.
Every candidate was able to draw the diagram correctly and most of them correctly
evaluated the required probabilities.
Some candidates were not clear about the distinction between a stationary
distribution and a limiting distribution.
(i)
Prices and salaries are notoriously non-stationary processes, having a tendency

to increase rather than a tendency to stay in the vicinity of some central value.
What is more, the increase is more likely to be geometric than linear. There is
some hope that {ln Pt}, which is equal to {ln (Pt/Pt-1}, may be a stationary
process, and similarly for S.
(ii)
If prices have recently increased, it is reasonable that workers will demand

salary increases; if salaries have recently increased, there is more money in
the economy, generating a tendency for prices to rise. The dependences are
reasonable.
(iii)
ln St a1 a 2 ln St -1 qS eS ,t
.
=
+ +
ln Pt b1 b2 ln Pt -1 qP eP ,t
This is a first-order VAR.

(iv)
Page 2
Sensitivity analysis comes after model verification. In model verification you

check that simulations of the process (with parameter values equal to the
values estimated from data) look similar to what has actually been observed.
For sensitivity analysis you check that this still works when the parameter
values used for the simulation are a little bit different from the estimates. The
purpose is to guard against the possibility that you have by chance used
parameter values with untypical properties.
Parts (i) and (ii) were done quite well. In part (i) a lot of people concentrated just on
the logs or just on the differences rather than both. In part (ii) quite a few people gave
answers such as "both are linked to inflation" rather than explaining why, or only
considered one way (eg why should salaries be related to prices, but not the other
way around).
Most people got part (iii). In part (iv) most people mentioned varying the parameters
slightly, but not many conveyed the idea of then studying the simulated output of the
model and assessing whether it still looked similar to real life.
(i)
E(Yt) = mt, Var(Yt) = s2t.
(ii)
E(Xnh) = n(2p - 1) D, Var(Xnh) = nD2 4p(1 - p).

t
t
Therefore E(Xt) = (2p - 1) D , Var(Xt) = 4p(1 - p)D2 .
h
h
(iii)
(a)
We require m = (2p -1)D/h and s2 = 4p(1 - p)D2/h.
(b)
This implies that p =
1 mh 2
2
2 2
1 + , D = s h + m h .
2
D
This applies only when 0 |h| < m-1. Small values of h should be used
because the random walk model converges to Brownian motion /
diffusion as h 0.
Parts (i) and (ii) were well answered maybe half the people successfully equated
the random walk moments with the Brownian motion ones although many
candidates failed to see the derivation of the conditions on p and D through to
completion.
The biggest problem was the last part: "h small" is on the whole a bit too vague to get
the full credit.
(i)
Let X = - (log U)/q. Then, for x > 0,

P(X > x) = P(log U < -qx) = exp(-qx).
Differentiating, f ( x ) = qe-qx , as required.
Page 3
Alternatively: use the inverse distribution function method.

x
F ( x) = q e -q y dy = 1 - e -q x . We need to invert this: set U = F(X) and express

0
X as a function of U. This gives X = - q -1 log (1-U).

(ii)
(a)
Use f(x) = q exp(-qx) as the base density. We need to find a constant

e-qx
C such that k (q)
C qe -qx for all x > 0.
1+ x
C = k(q)/q is the best that we can do.
The procedure is:
(b)
1.
Generate a value y from the density f(x) = q exp(-qx).
2.
Take another Uniform pseudo-random variable U2; if this is less

than g(y)/(C f(y)) [which is equal to 1/(1 + y)] then we accept the
value y, otherwise reject it and return to 1.
On average it takes C repetitions of steps 1 and 2 to generate a value.

Each such repetition requires two uniform pseudo-random variables.
So the answer is 2 k(q)/q.
Most candidates did well here, although a few people inverted the density function
instead of the distribution function.
Some experienced real difficulty in providing a clear statement of the algorithm for
Acceptance-Rejection method. A lot of people answered this part in abstract without
realising that they should use the density from part (i) as the base density, and hence
didn't get the marks for calculating C.
(i)
(a)
EW(t) = 0, Cov(W(s), W(t)) = 16Cov(B(ks), B(kt)) = 16k min(s, t).

Thus k = 1/16 is the required value.
To prove that a process is a BM, it is necessary to check the
covariance, not just the variance.
(b)
Page 4
In addition to the expectation and covariance function, we need to

show
that W is Normally distributed. Any linear transformation of a

Normal random variable is itself Normal.
that W has continuous sample paths. For small h, we have

W(t + h) - W(t) = - 4(B(kt + kh) - B(kt)), which clearly tends to 0
as h approaches 0.
*
(ii)
An alternative to proving continuity is to state that the increments

of W are independent of past values and are also Normally
distributed.
Consider exp[2B(t)-2t] with respect to the filtration Ft.

B(t) B(s) is independent of Ft and B(s) is Fs-measurable. Then
) (
E (e
E (e
E e 2 B (t ) | Fs = E e2[ B (t )- B ( s )]e 2 B ( s ) | Fs
= e2 B ( s )
2[ B (t )- B ( s )]
= e2 B ( s )
2[ B (t )- B ( s )]
| Fs
The increment B(t) B(s) has the normal distribution with mean 0 and
variance t - s, so the expectation of e2[B(t) B(s)] is equal to
M(2) = exp(2(t - s)), where M() is the moment generating function of the
N(0, t - s) distribution.
It follows that
E e 2 B (t )-2t | Fs = e2 B ( s )-2 s
and therefore e2[B(t)- t] is a martingale.
For part (i), in (a) many candidates just checked variance rather than the covariance,
but got 1/16 correct. Not so many people got part (b) generally people assumed
that (a) implied (b), rather than considering other properties of Brownian motion
Most candidates fared well on part (ii), identifying where necessary the mgf of a
normal random variable to successfully demonstrate the given process is a
martingale.
Page 5
(i)
{X(t)} is not Markov because, for example, P[Xt+1 = 2Xt = 3, Xt-1 = 2, ]

cannot be reduced to P[Xt+1 = 2Xt = 3].
(ii)
(a)
Define new states

3a = Level 3 this year following Level 2 last year
3b = Level 3 this year following Level 4 last year
(b)
The transition matrix is then

1
2
3a
4
3b
(c)
1
0.2
0.2
0
0
0.2
2
0.8
0
0.2
0
0
3a
0
0.8
0
0
0
4
0
0
0.8
0.8
0.8
3b
0
0
0
0.2
0
We have
p1 = 0.2p1 + 0.2p2 + 0.2p3b
p2 = 0.8p1 + 0.2p3a
p3a = 0.2p3b + 0.2p3a
p3b = 0.2p4
Applying the condition
p1 = 0.25p2 + 0.25p3b
p2 = 0.25(p3a + p3b)
p3a = 0.25p3b
pi = 1, we obtain the solution as
(p1, p2, p3a, p4, p3b) =
1
(21, 20, 16, 320, 64).
441
It follows that the long-run proportion of time spent in Level 3 is

(16 + 64)/441 = 80/441.
A large number of candidates fared well on this question. Some committed the mostly
harmless error of including too many (redundant) states. Most people had a
reasonable attempt at (ii)(c) too, although there were some errors of arithmetic or
matrix multiplication.
(i)
The generator matrix is

s
m
-s - m
-r - n n .
r
0
0
0
(ii)
Page 6
Let B and C be the levels of benefits and contributions, and define h

(respectively s) as the expected time spent by a policyholder in state H
(respectively S) before finally entering state D. For solvency the company

requires Ch - Bs 0.
The model allows h and s to be calculated. [The equations are
h=
(iii)
1
s
+
s,
s+m s+m
s=
1
r
+
h .]
r+n r+n
(a)
This might improve predictive power as far as the individual

policyholder is concerned, but in a large population of policyholders
with constant age profile it is not likely to make much difference.
(b)
Certain kinds of sickness are more likely to strike at particular times of

year, but they may more or less balance out. If there is a significant
increase in the incidence of sickness in one season, it should be helpful
to include this in the model.
(c)
Similarly to (a), the inclusion of duration dependence will significantly

improve the goodness of fit when only one policyholder is being
regarded, but in a large population of policyholders it is unlikely to
make a difference overall.
(iv)
If the population is split up into categories depending on age, duration of

illness and time of year, it is highly unlikely that there will be sufficient data to
estimate all the transition rates reliably.
(v)
This can be regarded a Time Series problem: we seek a test for the existence
of seasonal variation. One suggestion could be to find the best-fitting model
without seasonal variation and the best-fitting with seasonal variation, then to
compare the values of the Akaike Information Criterion.
Alternatively, if the number of policyholders is roughly constant from year to
year, one could use one-way Analysis of Variance to determine whether
Season has a significant effect on claims. If the number of policyholders
varies greatly from year to year, a two-way ANOVA with Year and Season as
explanatory variables could be fitted.
A test based on c2 would have to be carefully constructed. For example,

drawing up a contingency table with the quarters as the rows and with column
titles like Stayed healthy and Fell sick would be a reasonable approach.
Again done quite well generally.
Many candidates weren't clear that they needed to write a formula that characterises
the company's policy in part (ii).
Parts (iii) and (iv) done well most people seemed to get the general ideas and write
sensible comments, even if they didn't quite get full marks. Only a few candidates
Page 7
noticed that in a large stable population of policyholders the trainees suggestions

would not greatly improve the models predictive power.
In part (v) many candidates merely mention using the chi-squared test without
explicitly describing it. The question was carefully written to elicit responses which
showed whether the candidate understood what was to be tested.
(i)
State space for Xt is {0, 1, 2, 3}.
(ii)
(a)
Let T be the time taken for a cash dispenser to break down. From the
question:
P (T ( t , t + dt ) | T > t ) = a dt + o(dt )
or in other words
P (T ( t , t + dt )
P (T > t )
) = adt + o(dt )
If we set F(t) = P(T t), then

F ( t ) dt
1 - F (t )
= adt + o(dt )
and letting dt 0,
F (t )
1 - F (t )
=a
Integrate this to get - ln (1 - F ( t ) ) = a t + const

Since F(0) = 0, we can set const = 0, and hence F(t) = 1 e-t
That is, T has an exponential distribution with mean 1/.
(b)
(iii)
P(no breakdowns by time t) = (e-at)3 = e-3at. Thus the time until the
first breakdown is exponentially distributed with parameter 3a.
P(X(t + h) = m) = P(X(t) = m, X(t + h) = m) + P(X(t) = m - 1, X(t + h) = m)

+ P(X(t) = m + 1, X(t + h) = m) + o(h).
Thus in general
Pm (t + h) = Pm (t )[1 - ((3 - m)a + mb)h] + Pm-1 (t )(4 - m)ah1m>0

+ Pm+1 (t )(m + 1)bh1m<3 + o(h),
Page 8
where 1A is the indicator function of event A.

Rearranging and letting h 0 gives the required equations.
(iv)
First look at the LHS: if Pm has the given form, then

3
Pm' (t ) = q(t ) m-1 (1 - q(t ))2-m q '(t )[m(1 - q(t )) - (3 - m)q(t )]
m
On the other hand, the RHS is equal to

3
- q(t ) m (1 - q(t ))3- m [ mb + (3 - m)a ] +
m
3
3
m -1
4- m
m +1
2- m
(4 - m)a
+ (m + 1)b
q(t ) (1 - q(t ))
q(t ) (1 - q(t ))
1
1
m
m
+
3
= q(t )m-1 (1 - q(t )) 2-m -[mb + (3 - m)a ]q(1 - q) + ma (1 - q) 2 + (3 - m)bq2
m
3
= q(t )m-1 (1 - q(t )) 2-m (m - 3q) {a (1 - q) - bq}
m
Thus the LHS and RHS are equal as long as q' (t ) = a - (a + b)q(t ) .
Almost everyone got part (i). In part (ii), stating the time until first breakdown is
exponential in (a) is not sufficient: it was necessary to derive this from first principles.
Similarly, a reasonable number got the answer to (b) without deriving it properly.
The most successful strategy for part (iii) was writing the generator matrix and
getting the equations from there.
Only the exceptional candidate attempted (iv); the algebra was tough, but did not
involve any tricks.
(i)
We need to prove that E(Yn+1X1, X2, Xn) = Yn.

We have
E(Yn+1X1, X2, Xn) = E ((q / p ) Sn + X n +1X1 , X 2 ,... X n )
= (q / p ) Sn E ((q / p ) X n +1X1 , X 2 ,... X n )
Page 9
= (q / p) Sn ( p(q / p ) + q (q / p ) -1 + (1 - p - q )) = Yn ,
which shows that Yn is a martingale.
By taking expectations, we obtain that E(Yn+1) = E(Yn), and this in turn implies
that E(Yn) = E(Y0) = (q/p)m.
(ii)
(a)
It is clear that T is a stopping time with respect to this martingale, since

the event that the share has reached 0 or N by time n depends only on
X1, X2, , Xn and m.]
For n T, we have 0 Sn N,
Yn = (q / p ) Sn (q / p ) N if q > p, or Yn 1 if q p.
(b)
(iii)
The conditions of the optional stopping theorem are satisfied. It

follows that E(YT) = E(Y0) = (q/p)m.
m
N
0
q N
q
q
q
= E (YT ) = P( ST = N ) + P( ST = 0) = 1 - 1 - P( ST = N ).
p
p
p
p

Thus P( ST = N ) =
(iv)
(a)
1 - (q / p) m
1 - (q / p) N
so that P( ST = 0) =
(q / p)m - (q / p) N
1 - (q / p) N
We have
E(Zn+1X1, X2, , Xn) = E ( Sn2 + 2Sn X n+1 + X n2+1 - 2(n + 1) pX1 , X 2 ,..., X n )
= Sn2 - 2(n + 1) p + 2Sn E ( X n +1) + E ( X n2+1 ).
But E(Xn+1) = 0, E ( X n2+1 ) = p + q = 2p, and the last equation gives

E(Zn+1X1, X2, , Xn) = Sn2 - 2np = Z n , so that {Zn} is a martingale.
(b)
The conditions of the optional stopping theorem are not satisfied, since
Sn2 - 2np is not bounded below for 0 n T. But we can work with a
truncated stopping time TK = min(T, K), for which the conditions of the
OST are satisfied, then let K .
Applying the optional stopping theorem we get
E(ZT) = E(Z0) = m2.
Page 10
But E(ZT) = E ( ST2 ) - 2pE(T) = N2P(ST = N) - 2p E(T) and we are

given that P(ST = N) = m/N. Thus m2 = E(ZT) = Nm - 2pE(T), which
implies that E(T) = m(N - m)/(2p), as required.
Parts (i) and (iv)(a) were generally answered successfully.
Part (ii): in (a), the bounds on (q/p)Sn must be shown to apply for any p and q, not
just some; the main problem with (b) was that many did not mention the Optional
Stopping Theorem as justification for claiming that E(YT) = E(Y0).
Of the candidates who attempted (iv)(b), very few noticed that the stopping time did
not meet the requirements of the OST, so needs to be truncated.
10
(i)
The model can be written as

(1 - 1.7B + 0.4B2 + 0.3B3) Xt = (1 - 0.7B + 0.12B2) et.
The term in the brackets on the LHS above is divisible by 1 - B. We have
(1 - 1.7B + 0.4B2 + 0.3B3) = (1 - B)(1 - 0.7B - 0.3B2).
The last term is also divisible by 1 - B, giving
(1 - 1.7B + 0.4B2 + 0.3B3) = (1 - B)(1 - 0.7B - 0.3B2) = (1 - B)2(1 + 0.3B).
(ii)
The model can be identified as an ARIMA(1, 2, 2) process.
(iii)
{Xt : t 0} is clearly a non-stationary process, as the AR operator

(1 - B)2(1 + 0.3B) has two roots with modulus one; any ARIMA(p, d, q)
process with d > 0 is non-stationary.
(iv)
(1 + 0.3B)Wt = (1 - 0.7B + 0.12B2)et, which can be expressed as

Wt = (1 + 0.3B)-1(1 - 0.7B + 0.12B2)et.
Expanding the first term, Wt = (1 - 0.3B + 0.09B2 - )(1 - 0.7B + 0.12B2)et
= (1 - B + 0.42B2 - )et.
Therefore y0 = 1, y1 = -1, y2 = 0.42.
(v)
Just invert the previous representation: et = (1 - B + 0.42B2 - )-1 Wt = (1 +

B - 0.42B2 + B2 - ) Wt, so that p1 = -1, p2 = -0.58.
(vi)
None of the three processes is Markov; we know that if {Xt : t 0} is an

ARIMA(p, d, q) process with q > 0, then any finite collection (Xn, Xn-1, ,
Xn-m+1)T is non-Markov.
Page 11
(vii)
First transform the data, so that the differenced observations yt = (1 - B)xt may
be thought of as realisations of a stationary ARIMA(1, 1) model.
Estimate the parameters of the model (by Maximum Likelihood or Method of
Moments) and obtain forecasts for future values of y.
Transform these back to obtain forecasts for the x values.
Part (i) was done very well; of people with the wrong answer, most stopped after
taking out one (1-B) factor rather than trying for the 2nd. (ii) and (iii) were also well
answered, following on from (i).
In (iv) there were difficulties with the expansion of the denominator, meaning that few
candidates did well. Similar problems were encountered with part (v).
For the verification or otherwise of the Markov property in (vi) it would have helped
to write out {Yt} and {Zt} explicitly in terms of lagged terms.
Most candidates who attempted (vii) got at least part of the method, but few wrote
down all the main steps in the correct sequence.
Page 12
EXAMINATIONS

1.
booklet.
2.
supervisor.
3.
4.

question paper.
103S2003
A wheel is divided into 37 equal sections, labelled from 0 to 36. A ball is rolled
repeatedly around the wheel, landing in sections X1, X2, Let Mn be the highest
outcome achieved from the first n rolls, i.e. Mn = max1in Xi.
(i)
Show by general reasoning that Mn is a Markov chain.
[1]
(ii)
Derive the transition probabilities for Mn.
[2]
(iii)
Determine whether Mn is irreducible and aperiodic.
[1]
(iv)
Determine the equilibrium probability distribution of Mn.
(i)
Calculate the covariance between the values X(t), X(t + s) taken by a Poisson
process X(t) with constant rate l at the two times t and t + s, where s > 0. [2]
(ii)
Calculate the covariance between the values B(t), B(t + s) taken by a standard
Brownian motion B(t) at the two times t and t + s, where s > 0.
[1]
(iii)
A share price, St, is modelled as St = exp(Y(t)), where
[1]
[Total 5]
Y(t) = Y(0) + sB(t) + k(X1(t) - X2(t)).

Here B(t) is a standard Brownian motion, X1 and X2 are rate l Poisson
processes, independent of each other and of B(t), and s and k are constants.
(a)
(b)
(c)
Give a definition of a Lvy process.

Show that Y(t) is a Lvy process.
Calculate the covariance of Y(t) and Y(t + s) for s > 0.
[4]
[Total 7]
An ARIMA process X satisfies the recursion

X t = a1 X t -1 + a 2 X t -2 + et + bet -1 ,
where et is white noise with variance 2.
(i)
(a)
Write down a condition in terms of the roots of an equation for X to be

stationary.
(b)
Show, in the case where a2 = -0.5, that X is stationary as long as

a1< 1.5.
[4]
(ii)
103 S20032
Derive the spectral density function of X.
[3]
[Total 7]
Assume that the yield rate Yt of a certain stock may either take a high constant value
yu or a low constant value yd, during alternating random exponential periods with
respective intensities lu, ld.
(i)
Write down Kolmogorovs forward equation for the probability Pu,u(t) that Y
is in state yu at time t, given that it starts in state yu.
[1]
(ii)
Show that
Pu,u =
(iii)
ld
lu
e- ( lu +ld )t .
+
lu + l d lu + l d
Let Ut denote the total amount of time spent in state yu up until time t. Derive
and expression for E[UtY(0) = yu], the expected occupation time in state yu
by time t for the two-state continuous-time Markov chain starting in state yu.
Hint: Note that Ut =
result.
(iv)
[2]
1 if Ys = yu
and use the previous
I
ds
,
where
I
=
s
s
0
0 if Ys yu
[2]
t
Derive an expression for the expected total yield Xt of the stock by time t.
Hint: Xt = yuUt + yd(t - Ut).
[2]
[Total 7]
Let X1, X2, X3, be independent, identically distributed random variables with
P(X1 = +1) = p > and P(X1 = -1) = 1 - p < . Let S0 = 0, Sn = X1 + X2 + + Xn
for n 1 be the associated random walk.
(i)
State a necessary condition which must be satisfied by the constants q and c if

the process
Mn = eqSn -cn
is to be a martingale.
[1]
(ii)
Use the condition in (i) above to solve for q as a function of c.
[2]
(iii)
Let T1 be the first time that Sn hits 1, i.e. T1 = min{n : Sn = 1}. State the
conditions on c and on q(c) under which it is valid to use the Optional
Stopping Theorem to evaluate E ( M T1 ) .
[2]
(iv)
103 S20033
Derive the moment generating function E (e- cT1 ) for c > 0.
[3]
[Total 8]
PLEASE TURN OVER
(i)
Calculate the autocovariance function {gk : k 0} and the autocorrelation

function {rk : k 0} for the mth order Moving Average process
Xt = m +
1
(et + et-1 + + et-m),
m +1
where {et : t 0} is a sequence of uncorrelated, zero-mean random variables

with common variance se2 .
(ii)
[4]
Explain whether or not the process is invertible in the case where m = 2. [4]
[Total 8]
The data set plotted in Figure 1a represents the number of applications, xt, for travel
insurance received by an insurance companys web site, measured for a total of 60
consecutive months. Figure 1b displays the logarithm of the same data set, yt = ln(xt).
(i)
A statistician decides, on the basis of these plots, to fit a linear time series
model to yt rather than to xt. State, giving reasons for your answer, whether
you agree with this decision.
[2]
(ii)
Explain what is meant by the terms seasonal variation and linear trend.
Outline one method of compensating for seasonal variation and linear trend in
a data set which exhibits both.
[3]
(iii)
The data set zt is a seasonally adjusted, detrended version of yt. Figures 2a and
2b below display respectively the sample autocorrelation function and sample
partial autocorrelation function of zt.
(iv)
103 S20034
(a)
Explain what feature of the graphs enables you to conclude that it is

reasonable to fit a stationary model to the data.
(b)
Suggest values of p, d and q such that an ARIMA(p, d, q) model is

likely to provide a good fit to the data set zt. Give reasons for your
suggestion.
[3]
Explain how to produce a forecast x60 (1) for the value of x61 given the BoxJenkins forecast z60 (1) for the value of z61.
[2]
[Total 10]
Volume of travel insurance, x t
10
20
30
40
50
60
10
20
Month, t
30
40
50
Month, t
Figure 1a
Figure 1b
Autocorrelation function of z t
0.6
ACF, r k
0.4
0.2
0
-0.2 0
10
11
-0.4
-0.6
Lag, k
Figure 2a
Partial autocorrelation function of z t
PACF
Log(Volume), y t = ln(x t )
0.3
0.2
0.1
0
-0.1 0
-0.2
-0.3
-0.4
-0.5
10
11
Lag, k
Figure 2b
103 S20035
PLEASE TURN OVER
60
A no-claims discount scheme has six classes of discount numbered from 0 (no
discount) to 5 (maximum discount). A claim-free year results in a move to the next
higher discount status (or in the retention of the maximum discount status); similarly a
year with one or more claims results in a move to the next lower discount class (or in
the retention of the no-discount status).
(i)
Assuming that the probability of a claim-free period is a, where 0 < a < ,

write down the transition graph and the transition matrix of the discrete-time
Markov chain which models this scheme.
[3]
(ii)
(a)
Write down the equations obeyed by the stationary probability

distribution p = (p0, p1, , p5) of the Markov chain.
(b)
Solve successively for p1, p2, p3, , p5 in terms of p0.
(c)
Derive an expression, in terms of a only, for the value of p0.

[5]
(iii)
Explain how this analysis can be used to help the insurance company to set its
premium levels.
[2]
[Total 10]
A model of mortality after retirement suggests that the age at death, T, of an

individual aged 65 at retirement is a random variable satisfying
t
P (T > t ) = exp - m ( x)dx ,
65
where m(x), the force of mortality, is given by m(x) = a + bx, and a and b are positive
parameters whose values are known.
(i)
(ii)
(iii)
103 S20036
Determine a formula involving a uniform pseudo-random variable U for

simulating the age at death, T, of a single individual currently aged 65.
[5]
As part of a simulation exercise, a trainee is required to consider a sample of

10 individuals now aged 65 and to generate a single observation of Y, the
number of individuals in that sample who will still be alive at age 75.
(a)
Describe a method for doing this which does not involve 10 repetitions
of the method in (i).
(b)
Explain the main advantages of the method in (a) when a large number
of simulated copies of the random variable Y are needed.
[4]
The trainee is asked to investigate the effect of a reduction in the mortality rate
by repeating the whole simulation for different values of a and b. State with
reasons whether the trainee should use the same sequence of pseudo-random
numbers as before or whether a different sequence would be preferable.
[1]
[Total 10]
10
(i)
Define a geometric Brownian motion and write down a stochastic differential

equation which it satisfies.
[2]
(ii)
Using the substitution Yt = eat ( X t - c ) , or otherwise, find a solution for the

stochastic differential equation
dX t = - a ( X t - c)dt + sdBt
with the initial condition X0 = x0, where Bt is standard Brownian motion.
[5]
(iii)
Write down a condition satisfied by the stationary density function, p(x), of a

diffusion process, assuming that such a stationary density exists.
[2]
(iv)
Verify that the stationary distribution of the diffusion process in (i) is Normal,
and identify the mean and variance.
[2]
(v)
When the equation in (iii) is applied to the geometric Brownian motion, the
general solution is found to be
A
p ( x) = + Bx k ,
x
2
where k = -2 + 2m/s and A and B are arbitrary constants. Show that there is
no probability density function p(x) which solves the equation in (iii) in the
case of the geometric Brownian motion and explain the implication of this. [2]
[Total 13]
103 S20037
PLEASE TURN OVER
11
A continuous-time process with three states is observed from time 0 up until the time
of the 20th transition. The results may be summarised as follows:
State, i
No. of visits
to state i
Minutes spent
in state i
1
2
3
8
4
8
48
160
240
No. of transitions
from state i to:
State 1 State 2 State 3
1
7
3
1
5
3
-
(i)
Describe the stages of model fitting and model verification in the modelling
process.
[2]
(ii)
Suppose that a Markov jump process model is to be fitted to the data set
above. List all the parameters of the model and discuss the assumptions made
when such a model is fitted to a data set.
[4]
(iii)
Estimate the parameters of the model in (ii) above and write down the
estimated generator matrix.
[4]
(iv)
Suggest one test which could be applied as part of the model verification
process. State the null hypothesis, H0, identify the test statistic and name its
distribution under H0.
[2]
(v)
The 20th transition of the observed process takes it into state 1. Use the
estimated parameter values to give point estimates of the times until the 21st
and 22nd transitions.
[3]
[Total 15]
103 S20038
REPORT OF THE BOARD OF EXAMINERS

September 2003
Subject 103
Stochastic Modelling
EXAMINERS REPORT
EXAMINATIONS
September 2003
Subject 103
EXAMINERS REPORT
Subject 103 (Stochastic Modelling)
September 2003
Examiners Report
(i)
Suppose that Mn = i. Then the value of Mn+1 depends entirely on the outcome
of the next spin of the roulette wheel (this being independent of Mn) and the
value of Mn. Hence Mn is a Markov chain.
(ii)
Let Xn be the winning number from the nth spin of the roulette wheel. Then
P(Xn = i) = 1/37 for all i = 0, 1, 2, .,36.
If j > i then P(Mn+1 = j | Mn = i) = P(Xn+1 = j) = 1/37;
If j < i then P(Mn+1 = j | Mn = i) = 0;
If j = i then P(Mn+1 = j | Mn = i) = P(Xn+1 i) = (i + 1)/37.
1
i.e. pij =
37
(i 1)
37
if
if
if
(iii)
M is aperiodic, as it can stay in the same state with positive probability. It is

not, however, irreducible, since it is not possible to return to state j from any
state k > j.
(iv)
State 36 is absorbing, so the only stationary distribution, which is also the

limiting distribution, is 36 = 1, i = 0 for all other i.
The general reasoning type answers required for parts (i), (iii) and (iv) were quite well done in
general. There were some difficulties with part (ii) However candidates fared less well on part (ii),
and in many cases candidates struggling on part (ii) then failed to attempt the later parts.
(i)
The Poisson process and the standard Brownian motion both possess the
independent increments property.
Cov(X(t), X(t + s)) = Cov(X(t), X(t)) + Cov(X(t), X(t + s)
the independent increments property.
X(t)) = t + 0, by
(ii)
Cov(B(t), B(t + s)) = Cov(B(t), B(t)) + Cov(B(t), B(t + s)
B(t) = t.
(iii)
(a)
A Lvy process is a continuous-time process with stationary,

independent increments. Alternatively, a Lvy process can be defined
as a sum of three (indpendent) components: a constant drift, a multiple
of Brownian motion and a purely discontinuous random component
such as a compound Poisson process.
(b)
The increments of Y are the weighted sum of the increments of B, X1

and X2, so are stationary and independent.
Page 3
(c)
September 2003
Examiners Report
Cov(Y(t), Y(t + s)) = 2Cov(B(t), B(t), B(t + s)) + 2Cov(X1(t),

X1(t + s)) + 2Cov(X2(t), X2(t + s)) = ( 2 + 2 2 )t. All other terms
vanish by independence.
The question demanded straightforward manipulation of the independent increments

property and the covariance function.
(i)
(a)
The condition for X to be stationary is that the roots of the equation

1
1z
2z
=0
should lie outside the unit circle.
(b)
2
1
The roots are
In the given instance these are

If
2
1
2.
2
1
2 1
> 2 then we require that

2
1
2 1 and
which is equivalent to
2
1
implying that
1.5.
If, on the other hand,

2
z =
2
1
(2
2
1
< 2, then the roots are imaginary and satisfy
2
1)
=2
so that the condition is automatically satisfied.

(ii)
The spectral density satisfies

H1 ( ) f X
= H 2 ( ) f e ( ),
2
where H1 is the transfer function associated with (1
1B
2B ), H2 the
transfer function associated with (1 + B), fX( ) is the spectral density of X
and fe( ) is the spectral density of e.
Page 4
September 2003
Examiners Report
We have
H1( ) = |1
1e
2e
2i |2,
H2( ) = |1 + ei |2 and fe( ) =
2/(2
).
Therefore
2
fX
=
2
2
1
2
2
2 cos( )
cos
1 2
2 cos
There was good understanding of the ARIMA process, which meant that candidates
successfully derived the quadratic equation in part (i), though some were let down by their
knowledge of complex numbers. Part (ii) was a straightforward application of the transfer
function: the fact that marks were on average slightly lower seems to indicate that it had not
been learned especially well.
(i)
Pu ,u (t )
u Pu ,u (t )
d Pu ,d (t ) .
(ii)
Note that Pu ,d (t ) 1 Pu ,u (t ) . Therefore we have
Pu ,u (t )
d ) Pu ,u (t ) ,
implying that
d (
e
dt
d )t
Pu ,u (t )
de
d )t
Together with the boundary condition Pu,u(0) = 1, this gives the required
solution.
(iii)
E U t | Y0
yu
E I s | Y0
yu ds
P Ys
yu | Y0
yu ds
P ( s ) ds .
0 u ,u
Applying the previous part, this is equal to

d
u
(iv)
t
d
E X t | Y0
d)
1 e
d )t
yu = yd t ( yu
yd ) E U t | Y0
yd
( yu
yd )
d
t
u
u ( yu
yu
yd )
d)
(1 e
d )t
).
Page 5
September 2003
Examiners Report
In general candidates were able to score well on the first two parts, although a common mistake here
was to have the exponential parameters d and u transposed in the Kolmogorov Forward Equations.
In such cases appropriate credit was given for valid attempts at part (ii).
Sn cn
(i)
Mn = e
will be a martingale if E (e
(ii)
This will happen if ec = Ee
Xn
= pe
Sn
c ( n 1)
Fn ) = e
qe , i.e. if pe2
Sn cn
+c
+ q = 0.
Therefore
ec
= ln
(iii)
e 2c 4 pq
.
2p
For the OST to hold we require that M is bounded or T1 is bounded or M n
T1
is bounded.
In this instance, if c 0 and
0 then Mn < e for all n T1. But if c < 0 or
0 then there is no such upper bound and it is not safe to assume that the
OST can be applied.
(iv)
When c > 0, one root for is positive, the other negative, since
pe2 e +c + q < 0 when = 0. We need the positive root.
Applying the OST, E(MT) = M0 = 1 as long as c
that
1 = E (e
ST1 cT1
) = e E (e
cT1
0 and
0. This implies
).
Thus
E (e
cT1
) =e
2p
ec
e 2c 4 pq
ec
e 2c 4 pq
.
2q
In a number of cases, candidates covered some of part (ii) under part (i)
credit was given in these
cases. In general candidates were able to score well on the bookwork required for part (iii) although
candidates were less successful in tackling the final part.
Page 6
(i)
September 2003
Examiners Report
We have
= Cov(Xt, Xt k) =
1
(m 1) 2
(m 1)
Cov(et r , et
k r
k r ).
r 0r 0
0
m k 1
( m 1)2
et
r 0
For 0 k m, there are exactly (m

covariance terms equals e2 . Thus
et r ,
r 0
Clearly if k > m, all terms are zero, so that
Cov
2
2
e
= 0.
k + 1) non-zero terms, and each of these
0,1, 2,..., m.
The autocorrelation function is
(ii)
m k 1
m 1
k 1, 2,..., m
For the process to be invertible, we require that the roots of the characteristic
equation should be greater than 1 in absolute value.
We can rewrite the MA model with the aid of the backward shift operator B as
follows:
Xt
1
(1 + B + B2) et.
3
The roots of the characteristic equation

1 + B + B2 = 0
are B = + i 3 , B =
i 3 .
In both cases B = 1. Thus the process is not invertible.

In general candidates made reasonable attempts along the right lines, although this did not always
result in the correct autocorrelation function being calculated. Where possible, some credit was
given for attempts at the second part based on an incorrect part (i).
Page 7
September 2003
Examiners Report
(i)
The original data are clearly subject to seasonal variation, and the size of the
seasonal fluctuations is increasing in line with the value of the underlying
quantity. This suggests that the seasonal variation is multiplicative rather than
additive, in which case taking the logarithm is the sensible thing to do. In
addition to this, a look at the plot of yt against time confirms that the variation
is much more regular.
(ii)
Seasonal variation is a predictable pattern of deterministic variation in the

mean of the process which is cyclic, i.e it repeats after a fixed number of time
periods, usually corresponding to a year of elapsed time.
A linear trend is a deterministic pattern of variation in the mean of the process
which is linearly dependent on the time variable, i.e. is of the form a + bt.
There are various possible answers. Possible methods include:
(iii)
(iv)
(a)
Estimate the trend by linear regression and remove it, then, for each
month, calculate the sample mean value for that particular month over
all five years. From every detrended observation subtract off the
appropriate seasonal mean to obtain seasonally adjusted data.
(b)
Remove the linear trend by differencing the data once, then remove the
seasonal variation by seasonal differencing. In other words,
zt = (1 B12 )(1 B ) yt .
(a)
The fact that the sample ACF is not near 1 for small lags is the most
obvious pointer to the stationarity of the adjusted data set.
(b)
The clue here is the highest lag for which the ACF or PACF is
significantly different from 0. Looking at the sample ACF we might
suggest that a MA(3) might fit, as the sample ACF is roughly zero for
k > 3. Similarly, a look at the sample PACF seems to indicate an
AR(3). But it might well be possible to find an ARMA(1,1) which
would fit adequately. In other words, d = 0 and either p = 0, q = 3 or
p = 3, q = 0 or p = 1, q = 1.
We have yt = mt + zt, where mt represents deterministic variation and zt is a

purely random component. The process of seasonal adjustment and
detrending has produced an estimate mt for mt which can be extrapolated into
the future. Thus we have y60 (1)
m61 z60 (1) , which in turn leads to
x60 (1) exp( y60 (1)).

Neither the discussion of how to deal with seasonal variation nor the practical part to do with model
identification was especially well tackled.
Page 8
(i)
September 2003
Examiners Report
Transition graph:
(Transition probabilities are
to the right and 1
to the left.)
The transition matrix is

1
1
P=
(ii)
(a)
0
0
(b)
The equations P =
0(1
)+
0 +
1 +
2 +
3 +
0
0
read
1(1
)
)
)
)
)
2(1
3(1
4(1
5(1
4 +
=
=
=
=
=
=
0
1
2
3
4
5
Discard last equation and solve first one in terms of

1(1
)=
=
1
2(1
0(1
(1
0:
))
)=
2
0
1
j
In general
(c)
Find
0.
by normalisation:
Page 9
September 2003
Examiners Report
6
n
j 0
1=
1
0
1
(iii)
5
j 0
For a consistent profit the company requires that min0<

P( ) is the expected profit when annual claim rate is .
0.5
P( ) > 0, where
5
j 0
Expected long-term annual income from one customer is
j Pj , where
Pj
is the premium payable in discount level j, and expected annual claims

= C(1
), so expected profit is
P( ) =
5
j 0
Pj
1
1
C (1
), where
Candidates showed a good understanding of Markov chains here, with many candidates achieving a
high score. For the final part of the question, a large number of candidates only gave a cursory
explanation and were unable to score the full marks.
(i)
Use the inverse distribution function method.

F(t) = 1
P(T > t), so that
F (t ) = 1 exp
t
65
(a bx)dx = 1 exp
0.5b(t 2 652 ) a (t 65) .
Rearranging,
b 2
t at
2
b 2
65 65a log(1 F (t )) = 0,
2
or, in other words,

1
F (u ) =
Page 10
a 2 (65b) 2 2b 65a log(1 u )

b
September 2003
Examiners Report
Since the simulated variable must be positive, the positive root is required.
The method, then, is to generate a pseudo-uniform random variable U in the
range [0,1] and to set
T=
(ii)
(a)
a 2 (65b) 2 2b 65a log(1 U )

b
Y is a Binomial random variable, with parameters n = 10 and

75
exp
(a bx)dx
exp( 700b 10a ).
65
Let G(y) be the distribution function of Y and let U be a single uniform

pseudo-random variable on [0,1]. Then set Y min y : G ( y ) U .
(b)
(iii)
Although this method requires a certain amount of computing time to

evaluate the distribution function G, this only has to be done once;
thereafter, only one value of U is needed to generate each value of Y,
as opposed to the ten values of U which are required in the other
method.
It is important to use the same sequence of pseudo-random numbers in each

case, otherwise we are not comparing like with like.
Many candidates correctly identified that the inverse transformation method was required, although
in a surprisingly high proportion of cases marks were lost because of errors in the algebra /
integration. The later parts of this question, dealing with the comparison of two methods of
generating discrete random variables, were in general less well done.
10
(i)
A geometric Brownian motion can be defined as a process

St S0 exp( t
Bt ) , where Bt is a standard Brownian motion.
It satisfies the SDE dSt
St dt
St dBt .
Alternatively, use the definition St
dSt
(ii)
1
2
St dt
S0 exp( t
Bt ) , which satisfies the SDE
St dBt .
We use the It formula (we must have the form where F is a function of t as
well as x):
dF t , X t = Ft t , X t
Let Yt = e t ( X t
c)
1
2
Fx t , X t
dt
Fx t , X t dX t
F t , X t . Then Y(0) = x0 and
Page 11
dYt = d e t ( X t
c)
e t[Xt
c]
September 2003
e t[ Xt
Examiners Report
e t dBt
c] dt
= e t dBt
t
e s dBs
Hence Y t = x0 c
s 0
and so X t = c e
( x0 c)
e s dBs .
s 0
(iii)
The required condition for the stationary density of the diffusion Y solving
dYt
(Yt )dt
(Yt )dBt , from the Core Reading, is
d
d2
[ (y) (y)] = 2 [ 2(y) (y)]
dy
dy
(iv)
Two ways to do this. From (i),
X t ~ N (c e
( x0 c),
t
0
2 (t s )
The limiting distribution, as t

always stationary.
ds ) = N c e
, is N(c,
( x c) ( x)
1
2
( x0 c),
1 e
2 t
/(2 )). A limiting distribution is
Alternatively, in this instance we have (x) =

condition given in (ii) is that
( x)
(x c), (x) = , so the
( x) .
The solution to this DE is
( x) const.exp
( x c)2
2
which is the density function of N(c,

(v)
Page 12
2/(2
)).
A function of the form Axb for all x > 0 cannot integrate to 1, no matter what
the values of A and b. This means that there is no stationary density function
for the geometric Brownian motion.
September 2003
Examiners Report
It would be surprising if there had been: it is well known that Brownian

motion is non-stationary, and therefore that geometric Brownian motion is
also a non-stationary process, so cannot possess a stationary density.
Marks were poor for Question 10. Candidates should note that the reference in part (iv) of the
question to "the equation in (i)" was incorrect and should have read "the equation in (ii)". Whilst at
least some candidates were still able to complete this part of the question, allowance was made in the
marking of scripts for this error.
It seemed that candidates had not committed to memory the formula for the equilibrium density of a
diffusion process, presumably because it had not been asked before.
11
(i)
Model fitting: this occurs after the family of model has been decided and
concerns the estimation of the values of parameters. The set of parameters to
be estimated is determined by the choice of model family.
Model verification: once the model has been fitted we need to check that the
fitted process resembles what has been observed. Generally we produce
simulations of the process, using the estimated parameter values, and compare
them with the observations.
(ii)
The parameters are the rate of leaving state i, i, for each i, and also the jumpchain transition probabilities, rij for j i, where rij is the conditional
probability that the next transition takes the chain to state j given that it is now
in state i. Alternatively, one may regard the parameters as being ij, where ii
=
i and, for j i, ij = i rij.
Assumptions of the Markov model are that the duration of holding time in
state i has exponential distribution with parameter determined only by i and is
independent of anything that happened before the current arrival in state i, and
that the destination of the next jump after leaving state i is independent of the
holding time in state i and of anything that happened before the chain arrived
in state i.
(iii)
1
i
is the average duration of each stay in state i. Thus
10 per hour,
1
40
per minute or 1.5 per hour,
1
30
1
6
per minute, or
per minute, or 2 per
hour.
r12
3
8
, r13
5
, r21
8
1
, r23
4
3
4
, r31
7
8
and r32
1
.
8
Thus the generator matrix, in units of hr 1, is
Page 13
10
3.75
September 2003
Examiners Report
6.25
0.375 1.5 1.125

1.75 0.25
2
(iv)
One test should test whether the holding times in each state are exponentially
distributed. If Ti,k denotes the kth holding time in state i, then the hypothesis is
that Ti,1, Ti,2, , Ti,ni is a sample from an exponential distribution with
parameter
i:
sort the observations into categories, calculate expected number
in each category and hence find the X2 statistic by summing (O E)2/E. This
should be compared with the critical value of the 2 distribution with m 2
d.f., where m is the number of categories.
(v)
Estimate of expected duration of a visit to state 1 is 6 mins, so this is the

estimated time until the 21st transition.
Estimate of expected time between 21st and 22nd transitions is
E(Time | transition is to 2) P(transition is to 2) + E(Time | to 3) P(to 3)
= 40 (3/8) + 30 (5/8) = 33.75 mins.
Attempts to fit the model to the observed data were generally sensible and encouraging,
although in a minority of cases the calculation of the parameter estimates betrayed evidence
of some confusion. There was a tendency to be less successful as the question continued,
with the result that attempts at the final part were of a noticeably lower standard.
Page 14
This document was created with Win2PDF available at http://www.daneprairie.com.

The unregistered version of Win2PDF is for evaluation or non-commercial use only.
EXAMINATIONS
20 April 2004 (pm)
Subject 103

1.
booklet.
2.
supervisor.
3.
4.

question paper.
103
A2004
The stochastic differential equation for geometric Brownian motion is given by

dXt =
Xt dt +
Xt dBt
where Bt is a standard Brownian motion. Show that the solution to the differential
equation which also satisfies X t0 x is
Xt
for t
x exp
Bt
1
2
Bt0
t t0
t 0.
[4]
A Box-Jenkins model-fitting procedure suggests that the best fitting model for a set of
normalised share price data x1,
, xn is ARMA(1,2), with equation
Xt
0.63 X t
where {e1, e2,

variance 2.
et
0.45et
0.34et
2,
} is a sequence of uncorrelated, zero-mean random variables with
(i)
Determine whether the model is stationary and invertible.
(ii)
Calculate 0, 1, 2, the autocovariance function of the fitted model at lags 0, 1,

and 2, in terms of 2.
[4]
[Total 6]
103 A2004
[2]
The number, N(t), of members of a pension scheme who are receiving benefits at time
t, is subject to change of two kinds:
it increases by 1 when an active member reaches retirement age
it decreases by 1 when a retired member dies
Assume that retirements occur according to a Poisson process with rate and that
each retired member, independently, has a probability dt of dying within the time
interval (t, t + dt).
(i)
Explain why, under these assumptions, N(t) is a Markov jump process.
[1]
(ii)
Write down the transition rates of N(t).
[1]
(iii)
Using the notation pn(t) = P(N(t) = n), obtain a differential equation satisfied
by pn(t).
[2]
(iv)
Verify that one solution of the equation in (iii) is given by

n
1
pn (t ) = e
n!
(v)
n = 0,1,...
[2]
State what conclusions can be drawn from (iv).
[1]
[Total 7]
The value, St, of a stock exchange index is observed at hourly intervals over the
course of a week, generating observations s1, s2, , sn.
Three different models for St are being considered:
(i)
(a)
St = S 0
(b)
St = S 0 e
(c)
St
( St 1 ) et , where {et: t = 1, 2,
uncorrelated, zero-mean random variables.
Bt , where Bt is a standard Brownian motion.

Bt
, where Bt is as in (a).
} is a sequence of
Outline the processes of model fitting and model validation, including a

description of the role of simulation.
[3]
(ii)
For EACH of the three models above, outline the steps required during the
process of fitting the model.
[5]
(iii)
Choose ONE of the above models and describe TWO tests you would apply as
part of the model validation process.
[2]
[Total 10]
103 A2004
PLEASE TURN OVER
(i)
(a)
Write down the joint density function of B() and B(1), where
{B(t): t 0} is a standard Brownian motion.
(b)
Demonstrate that the conditional distribution of B() given that

B(1) = b is Normal with mean b and variance .
[5]
(ii)
James runs around a running track at the same time as an electric hare which
is programmed to go around in exactly one minute. Let Y(t) denote the
distance (in metres) by which James is ahead of the hare t minutes after the
start and suppose that Y(t) can be modelled as Y(t) = 5 B(t), where B(t) is as
in (i).
(a)
If James is leading the hare by 5 metres at time t = 0.5, determine the

probability that James eventually finishes the race ahead of the hare.
(b)
If James finishes the run 5 metres ahead of the hare, determine the
probability that he was ahead of the hare at time t = 0.5.
[5]
[Total 10]
(i)
Describe a method for generating a sequence U1 U 2

of pseudo-random
numbers uniformly distributed in the interval (0 1) and from this a pseudorandom sequence of numbers uniformly distributed over an arbitrary
range (a b) , where a b .
[3]
(ii)
Explain the main advantage of using pseudo-random, as opposed to purely

random numbers, for testing the suitability of a model.
[1]
(iii)
Given a sequence {U n n 1 2 } as in (i), explain how you would simulate

(a)
an observation from the Pareto density

f ( x) =
(b)
a
(1 x)a
103 A2004
a 1x 0
a discrete random variable X with probability function given by
P( X = i) =
(iv)
1
n 2
for i 1 2
P( X = 0) =
2
n 2
[4]
The Pareto distribution in (iii) is often referred to as a heavy-tailed

distribution . Explain the use of this term and discuss under what
circumstances it might offer a suitable model for modelling sizes of claims
arriving at an insurance company.
[2]
[Total 10]
(i)
Suppose that {Xt : t 0} is a time-homogeneous Markov jump process with

generator matrix A and transition matrix P(t). Let be a (column) vector of
probabilities such that TA = 0T, where 0 is a (column) vector whose
components are all equal to zero.
(a)
Prove that
d
dt
(b)
Explain why
P(t ) = 0T .
is known as the stationary distribution for X.
[4]
(ii)
A telephone call centre receives calls from customers at an average rate of 0.5
per minute. Each call has a random duration which is exponentially
distributed with mean 3 minutes, independently of the number or duration of
any other calls. Two operators are assigned to handle calls. If a call arrives
when both operators are busy, the call is put on hold unless there are already
two calls on hold, in which case the new call is lost. When a call ends, one of
the calls on hold is immediately put through to the newly free operator.
(a)
Identify the five states which are required if this system is to be

modelled as a Markov jump process.
(b)
Draw the transition diagram for this system.
(c)
Write down the generator matrix for the process.
(d)
Evaluate the stationary distribution of the system.

[8]
[Total 12]
103 A2004
PLEASE TURN OVER
A restorer of historic sites travels between Bath, Warwick, Stratford and Caernarvon.
Having arrived at a site, the restorer stays there for a random number of days, then
moves on to one of the others. The restorer s diary for the last two months shows the
following time spent in each place:
Bath (5 days), Warwick (3 days), Caernarvon (7 days), Bath (3 days),
Stratford (7 days), Warwick (5 days), Bath (3 days), Stratford (3 days)
Caernarvon (8 days), Bath (2 days)
A researcher suggests that the path taken by the restorer forms a Markov chain with
state space {Bath, Caernarvon, Stratford, Warwick}.
(i)
Suppose {Xt: t = 0, 1, 2, } is a Markov chain with transition matrix P and let

Di,n denote the duration of the nth visit to state i, that is, the number of
consecutive steps spent by the chain in state i. Show that Di,n is a geometric
random variable and is independent of Di,m for m < n.
[3]
(ii)
Discuss, in the light of the data, whether it is likely that this suggestion is
satisfied.
[2]
(iii)
Assuming that the researcher s suggestion is correct, write down an estimate

of the transition matrix P of the Markov chain.
[2]
(iv)
State, giving reasons for your answers, whether the Markov chain model is
(a) irreducible, and (b) aperiodic.
[2]
(v)
The restorer arrives in Warwick on day t0. Use the estimated transition matrix
to calculate the probability that he is in Warwick on day t0 + 3.
[3]
[Total 12]
103 A2004
A water company is developing a time series model to model the supply of water, Xt,
in its reservoirs at the end of month t. The model takes the form
Xt = Xt
Rt
Dt ,
where Rt represents rainfall in month t and Dt represents demand in month t. Rt and

Dt are themselves modelled by
Rt =
Dt =
( Rt
t 1)
et
Rt
where , , are positive constants, t is a deterministic function of time and

{et: t = 1, 2, } is a sequence of independent Normal random variables with mean 0
and variance 2.
(i)
(a)
Comment on whether it is reasonable to assume that demand is

negatively related to rainfall.
It is suggested that t should be cyclical with period 12, i.e. that
t+12 = 12 for all t. Comment on whether this suggestion is sensible.
[2]
(ii)
State, with reasons, whether Rt is stationary.
(iii)
(a)
Suppose that R0 is Normally distributed with mean 0 and variance 2R .

Show that Rt is also Normally distributed for each t > 0 and derive an
expression for its mean.
(b)
Calculate a value of
2
R
[1]
which ensures that Var(Rt) =
2
R
for all t. [4]
(iv)
Given a sequence of observations R1,

of data, suggest simple estimators for
, R120, D1, , D120 covering 10 years

[4]
t (1 t 12), , and .
(v)
Assuming that the values of the parameters are known exactly, and that R120,
D120 and X120 are known, derive an estimate x120 (1) for the supply of water at
the end of month 121 and give an expression for Var( X121 x120 (1)) .
[3]
[Total 14]
103 A2004
PLEASE TURN OVER
10
An insurance company has an initial capital of u and receives premium income at

constant rate c per unit time. Claims, for one unit each, occur according to a Poisson
process {Nt, t 0} with rate . The company s surplus at time t is defined as
St = u ct Nt .
(i)
A Lvy process is defined as a stochastic process with stationary, independent

increments. Verify that St satisfies this definition.
[2]
(ii)
For s > 0, write down the conditional distribution of Nt+s
(iii)
Define Yt = e
(iv)
St
Nt given Nt.
[1]
(a)
Derive an equation which must be satisfied by

martingale.
(b)
Show, by means of a sketch or otherwise, that the equation derived in

(a) has a negative solution as long as c > and a positive solution
if c < .
(c)
Comment on whether the condition c >

practice.
Let T = inf{t
0 : St
0 or St
if Yt is to be a
is likely to be satisfied in
[6]
K } , where K is a positive constant and assume
that c > .
(a)
Explain why either ST = K or 1
(b)
Verify that the conditions of the optional stopping theorem apply to the
stopped martingale Ymin(t ,T ) .
(c)
Define to be the probability that S goes below zero before it hits K,

that is, the probability that ST < 0. Prove that
E (YT )
(d)
(1
)e
ST < 0.
Derive a lower bound for using the result in (iv)(c) and state the
limit of this lower bound as K
.
[6]
[Total 15]
END OF PAPER
103 A2004
EXAMINATIONS
April 2004
Subject 103
EXAMINERS REPORT
April 2004
Examiners Report
First method (direct method):

Let f(Xt) = ln Xt . Using It s Lemma we have
df
X t dt
1
Xt
Xt
1
2
1
2
X t dBt
1.
2
X t2
dt
f (
X t2 dt
X t2 )dt
1
X t dBt
Xt
dBt
Integrating from t0 to t we have

ln X t
1
2
ln X t0
t t0
Bt
Bt0
and using the initial condition we get
Xt
x exp
Bt
1
2
Bt0
t t0
as required.
Alternative method (working backwards from the solution)

Let f (t , Bt )
f
t
df t
1
2
ft dt
x exp
1
2
f t dBt
Bt
2
Bt0
1
2
t t0
. Then
2
f
f
2
f,
f,
f , so that
2
B
B
1 2
f t dt
f t dt
ft dBt
2
This implies that Xt = f(t,Bt) satisfies the stochastic differential equation.

We need to verify that the initial condition holds for this solution. (It clearly does, but
the check needs to be performed.)
Considering that this stochastic differential equation is the most frequently used of all,
this question was on average very poorly answered.
(i)
The solution to 1 0.63 z

the model is stationary.
0 is z = 1/0.63, which is greater than 1. Therefore
For invertibility, we should check that the roots of 1 0.45 z 0.34 z 2 0 are
outside the unit circle. They are ( 0.45 1.25)/( 2*0.34) = 2.5 or 1.18, both
OK.
(ii)
Let
= Cov(Xt, et),
April 2004
= Cov(Xt, et 1),
Examiners Report
= Cov(Xt, et 2). Then
= 0.63 1 + 0 + 0.45 1 0.34 2

1 = 0.63 0 + 0.45 0 0.34 1
2 = 0.63 1 0.34 0
2
(this may be regarded as obvious and not stated explicitly)
0=
2 = 1.08 2
1 = 0.63 0 + 0.45
2 = 0.3404 2
2 = 0.63 1 0.34
0
An alternative expression for

0
Var(0.63 X t
0.632
et
0.45et
(1 0.45 2
0.34 2 )
may be obtained using

0.34 et 2 )
2
2 0.63[0.45
which removes the need to calculate
0.34
2.
Having derived the equations, we need to solve them.

= 0.63 0 + 0.0828 2
2
0 = 0.63(0.63 0 + 0.0828 ) + 1.370
1
= 0.3969
+ 1.422 2,
implying that
0
= 2.358 2,
= 1.5684 2,
= 0.6481 2.
Candidates were often unsure of the procedure required to derive the equations for the k but
did rather better at solving them. In particular, many candidates did not take correct
account of the relationship between Xt and past values of et. Marks were awarded for correct
methodology when deriving the solutions, even if the equations being solved were not the
right ones.
(i)
(ii)
It is a jump process because it remains in one state for a period of time and
then jumps to another state, or alternatively because it is a continuous-time
process with a discrete state space.
Given the entire past history of the process, the probability of a member
retiring and beginning to receive benefits in the next dt interval is dt, i.e.
independent of the past. The same applies to the probability of death between
times t and t + dt: it is dependent only on the state (the number of retired
members) at time t, and not on anything which happened before that. Thus the
Markov property holds.
n,n 1
n,n 1
Page 3
April 2004
Examiners Report
(iii)
From the Kolomogorov Forward equations we have that

dpn
( n,n 1
n,n 1 ) pn
n 1,n pn 1
n 1,n pn 1
,.
dt
( n ) pn (n 1) pn 1 pn 1
(iv)
The suggested form of pn does not depend on t, so its derivative is zero. The
RHS is
n
(
where
(v)
n )e
n!
n 1
(n 1) e
(n 1)!
n 1
(n 1)!
0,
= / .
The implication is that Poisson( ) is a stationary distribution for the Markov

jump process. (In this case it is also a limiting distribution, but that is not
deducible from the Core Reading.)
Very few candidates suggested that the rate at which deaths take place should be
proportional to the number of retired scheme members.
In (iv) examiners awarded marks for correct attempts to carry out the verification procedure
even when the differential equation in (iii) was wrong.
In (v), terms such as limiting distribution or equilibrium distribution were given full
credit.
(i)
Model fitting involves first choosing a suitable family of models, then a

suitable state space and finally estimating the values of the model parameters
to isolate a single member of the family.
Having performed the model fitting process, we have the best-fitting model
from the given family. Model validation involves checking whether this is an
adequate explanation of the observations.
Simulation can be used in model validation to test that the paths typically
followed by a simulated version of the process are broadly similar to the paths
observed in practice.
(ii)
(a)
Under the given model, the successive increments
St form a
sequence of i.i.d. N( , 2) random variables, so the parameters can be

estimated using the sample mean and sample variance of the observed
increments.
(b)
The first thing to do is to take the log of the observations, obtaining x1,
, xn say. Then apply the same technique as in (a).
(c)
April 2004
Examiners Report
This is a time series model, so a time series technique (such as BoxJenkins) would be appropriate.
Parameters are estimated by Method of Moments or Maximum
Likelihood, or simply by applying a computer package to do the
estimation for you.
(iii)
Any of the tests described here are equally valid.

(a)
The model implies that the increments are Normally distributed, so a

standard test of normality, such as the Anderson-Darling test, the
Kolmogorov-Smirnov test or a chi-squared goodness-of-fit test can be
employed to test this.
Another testable property of the model is independence of increments:
it is possible to investigate the lag-1 sample ACF and determine
whether it is significantly different from zero, or use the Ljung-Box
chi-square statistics to perform a portmanteau test on all lags of the
sample ACF simultaneously.
A similar test is to inspect the sample ACF and PACF of the residuals
to see whether the observed values fall outside the 5% significance
band.
We could test whether there was a relationship between the increment
st and st; this is a test for homoscedasticity, though in the situation
described in the question it is unlikely to provide useful results.
A turning points test, a run test or a sign change test could also be
used, in each case comparing a calculated statistic with tables of a
reference distribution. These are tests for serial independence (usually
applied to a sequence of residuals).
(b)
Apply the same tests as in (a) to the logarithm of the data.
(c)
Tests can be applied to the sequence of residuals arising from the

fitting process. Since, if the model is accurate, these should form a
sequence of uncorrelated Normal observations, the same tests as in (a)
can be applied.
A procedure which might be applied to the model in part (c) but not to
the models in (a) or (b) is to inspect the sample partial ACF of the
increment process St; if the process really is a first-order
autoregression as stated in the model, then the sample PACF should
have no significant values at lags higher than 1.
This question differentiated well between candidates, with some very good answers indicating
that the candidates understood the principles and practice of modelling, and others
highlighting deficiencies of understanding.
Page 5
April 2004
Examiners Report
(i)
(a)
If X = B() and Y = B(1), then the marginal distribution of X is N(0,0.5)
and the conditional distribution of Y given X = x is N(x, 0.5).
This means that f X ,Y ( x, y ) is given by
x2
1
.
exp
2(0.5)
2 (0.5)
1
exp
2 (0.5)
(b)
( y x)2
2(0.5)
exp( ( y 2 2 yx 2 x 2 ))
The conditional density function of X given Y is

f X |Y ( x | b)
exp
f X ,Y ( x, b)
fY (b)
exp( (b 2 2bx 2 x 2 ))
1
exp b 2
2
2( x b) 2
This is the density function of the Normal distribution with mean b

and variance .
(ii)
(a)
Here we seek P(Y(1) > 0 | Y() = 5) = P(B(1) > 0 | B() = 1)

=
P B( 12 )
= 1
(b)
2 B( 12 ) 0
2 = 0.9213.
We require P Y ( 12 ) 0 | Y (1) 5 = P B ( 12 ) 0 | B(1) 1

Using part (i)(b), this is 1
(1) 0.8413.
This question was designed in such a way that candidates could tackle part (ii) even if part (i)
had not worked out right, but many allowed themselves to become discouraged by difficulties
in the early part of the question and gave up.
(i)
Use a linear congruential generator (LCG). Specify three positive integers

a c m with m a m c and an initial value x0 , then generate a sequence of
x1 x2
xn in the range {0 1 2
m 1} by the recursive rule
xn
Then un
(axn
April 2004
c) (mod m)
n 12
Examiners Report
xn m is a sequence of pseudo-random numbers in [0 1] .
To generate a sequence over the range (a b) , a linear transformation

vn (b a )un a is needed. Thus the procedure is: generate a variable
U U (0 1) using a LCG as above; return V (b a)U a .
(ii)
The main advantage of using pseudo-random numbers is reproducibility.

When testing a model, this almost certainly depends on a number of
parameters and assumptions, so it is very often desirable to examine the
sensitivity of the model to these assumptions and the values of the parameters
used. It is thus necessary to reduce the random element to a minimum and
rerun the experiment using the same data .
(iii)
(a)
We use the inverse transform method. The cdf of the Pareto is

F ( x)
x
0
(1 x)
a 1
1
(1 x) a
so the inverse is
F 1 ( y ) (1 y )
1a
Thus the procedure is

Generate an observation U from the Uniform (0,1) distribution.
1a
1 (or, equally good, X

Return X (1 U )
observation from the Pareto.
(b)
1a
1 ) as an
Since the number of possible values for X is n 1 , we need to split

the interval (0 1) into n 1 subintervals, the length of which has to be
proportional to the corresponding probabilities. Given a uniform
variable U , one such procedure is:
For i 1 2 n , if (i 1) (n 2) U
If U n (n 2) , then set X 0 .
(iv)
i (n 2) , then set X
From (iii), we see that the tail of the Pareto distribution is 1 F ( x) (1 x) a ,

which decreases to zero much more slowly than e.g. the exponential or normal
distributions.
Alternatively, it is called fat-tailed because of the relatively high probability
of producing values which are a long way from the mean/median.
Page 7
April 2004
Examiners Report
In general, the Pareto offers a suitable choice for portfolios where there is a
non-negligible chance of very large claims, which is reasonable in various
forms of general insurance, in particular insurance that deals with natural
catastrophes (floods, earthquakes, hurricanes etc).
This question was generally well answered and caused few difficulties.
(i)
(a)
We will use the matrix form of the Kolmogorov DE, which states that
P '(t ) AP (t ) .
d T
T dP
T
It follows that
P(t )
AP(t ) 0T P(t ) 0T .
dt
dt
[The other Kolmogorov DE, P '(t ) P(t ) A , cannot be used to complete the proof]
(b)
If X0 is random with distribution , then the distribution of Xt is given

by TP(t).
The fact that the differential of this is equal to zero implies that TP(t)
= TP(0) = T for all t, in other words Xt has the same distribution for
all t.
For an irreducible Markov chain on a finite state space, the stationary
distribution is also the limiting (equilibrium, long-term, steady state)
distribution.
(ii)
(a)
The states can be labelled as 0, 1, 2:0, 2:1, 2:2, where 0 and 1 represent
the number of operators occupied and 2:j means that both operators are
occupied and j calls are on hold.
Candidates with a different collection of states can earn 0.5 marks, as
long as their states are states. For example, Call arrives is not a
state, but an event triggering a transition from one state to another.
(b)
1/2
0
1/2
1
1/3
1/2
2:0
2/3
1/2
2:1
2/3
2:2
2/3
The labels on the arcs represent the transition rates.

(c)
From the transition diagram, the generator matrix is as below.
1/ 2
A
(d)
April 2004
1/ 2
Examiners Report
0
0
1/ 3
5/6
1/ 2
2/3
7/6
1/ 2
2/3
7/6
2/3
1/ 2
2/3
We have
1
2
5
6
7
6
7
6
1
3
1
2
2:0
2:1
1
2
1
2
2
3
2:0
2:0
2
3
2:0
2:1
2
3
2:1
2:2
2:2
To solve this we add the condition that the

Therefore the stationary distribution is
128
653
192
653
144
653
108
653
3
2
9
0
8
27
0
32
81
0
128
i
must sum to 1.
81
653
(in decimal form this is

0.1960 0.2940 0.2205 0.1654 0.1240 )
Many candidates made a good attempt at part (i), though not many scored full marks. In part
(ii) a surprising number of candidates were unable to distinguish between states in which
the process stays for a certain length of time and events, which occur instantaneously and
trigger transitions from one state to another.
(i)
Suppose at time t X has just arrived in state i. The probability that X remains
in state i until time t + k and then leaves, giving a duration of k + 1 steps in
state i, is piik (1 pii ) . In other words, P ( Di ,n d ) piid 1 (1 pii ) , which is
the probability function of a geometric random variable.
The fact that Di,n is independent of previous durations follows from the
Markov property: What happens to X after arriving in state i is independent of
anything that happened before that moment.
(ii)
A sequence of successive observations of a geometric random variable is

likely to produce 1 more often than any other value. The durations of the
restorer s stays in the various locations are always at least 2 days and often
much more. There is an indication that the geometric distribution is
inappropriate for the data set provided.
Page 9
April 2004
Examiners Report
An alternative indication that the Markov property is dubious is the fact that
the restorer appears to return to Bath at regular intervals, i.e. regardless of the
time spent in each state, the path taken appears to lack randomness.
(iii)
pij
nij / ni , where nij is the number of transitions from state i to state j and
ni+ is the total number of transitions out of state i. For example, of the 8 days
spent in Warwick, one is followed by a trip to Caernarvon, one by a trip to
Bath and the remaining 6 are followed by another day in Warwick, so that nWC
= 1, nWB = 1 and nWW = 6. We therefore have (using the order B, C, S, W)
(iv)
9
2 1
0
12
12 12
2 13
0 0
15 15
1 8 1
0
10 10 10
1 1
6
0
8 8
8
The model is irreducible. Starting from Bath it is possible to visit Stratford,

Warwick, Caernarvon and return to Bath, completing the circuit.
It is also aperiodic, since pii > 0 for some (in fact for all) state i.
(v)
We need the entry of P3 corresponding to the row and column of Warwick

(4th row, 4th column in this case).
.5729 .0271 .2583 .1417

Now P 2
.2156 .7511 .0222 .0111

.0258 .1792 .6400 .1550
, so the required answer is
.2042 .2021 .0208 .5729
1
1
0
6
.1417
.0111
.1550
.5729 = 0.4488.
8
8
8
8
Each part of this question attracted some unexpected answers as well as some good
ones. The answers to part (i) were on the whole disappointing, but the final parts were
rather better answered in general.
(i)
April 2004
Examiners Report
(a)
In one sense it is reasonable, because in the absence of rainfall

gardeners might increase their demand for water. But the relationship
does not work in reverse: in particularly rainy weather the demand is
likely to be no less than in normally rainy weather.
(b)
Yes. It reflects the seasonal pattern of rainfall.
(ii)
No. It satisfies an equation which is explicitly dependent on t, so it will not be

stationary unless all the t are equal to one another. (Arguments based on
whether | | > 1 are only relevant if all the t are equal.)
(iii)
(a)
Normality: by induction. If Rt 1 and et are Normal, then Rt must be

Normal, as it is a linear combination of these two. Since R0 is Normal,
normality follows for all t.
E ( Rt )
[ E ( Rt 1 )
t 1] .
zero, showing that E ( Rt )
This can be iterated backwards to
An alternative approach is to solve explicitly for Rt and read off the

answers from there.
t 1
t
Rt
( R0
0)
et
k 0
Now we can see that Rt is a linear combination of Normal random

variables with some constants, hence Normally distributed, and that the
mean is t.
Var(Rt) is equal to 2 Var(Rt 1) + 2. If Rr has constant variance
then R2 = 2 R2 + 2, implying that R2 = 2 / (1 2).
(b)
(iv)
If instead the above expression for Rt is used, we can see that

t 1
1
1 2t 2
2
2t 2
2k 2
2
,
so
that
R
R
R
1 2t 1 2
k 0
t represents the mean rainfall in one particular month of the year. Estimate
this by calculating the average rainfall in that month over the 10-year period,
1
(rt rt 12 rt 24
rt 108 ) .
t
10
is best estimated using the lag-1 autocorrelation of R
120
(rt
)(rt
t 1
t 2
)
.
120
(rt
t 1
Noticing that Cov(Dt, Rt) =

Var(Rt), we suggest minus the ratio of the
sample covariance of D with R divided by the sample variance of the R, i.e.
Page 11
April 2004
Examiners Report
120
(rt
r )(dt
d)
t 1
, although in fact a better estimate would be obtained
120
(rt
r)
t 1
if account was taken of the variable mean of the Rt.

2
(Less appropriate is using the fact that Var( Dt )

2
sD
/ sR2 , since the sign of
is lost in this operation.)
Now we have an estimate for , we can use

(v)
X121
X120
D121
Var( Rt ) to suggest
R121 D121 . In addition, R121
Rt
Dt .
( R120
0)
e121 and
R121 .
Therefore X121
X120
implies that x120 (1)
(1
)[
X120
and Var( X121 x120 (1)) (1
(1
)2
)[
2
( R120
0 )]
( R120
(1
)e121 , which
0 )]
Most candidates made encouraging progress with this question, though part (iv) proved
difficult for most. In part (v), as well, the relationships between the variables were not
always well understood.
10
(i)
The increment St+s St is equal to cs (Nt+s Nt), which is independent of

anything that happened before time t and has the same distribution as
cs (Ns N0), using the independent increment property of the Poisson
process.
Alternatively, it is possible to use the Decomposition Theorem, to state that a
Lvy process is any sum of three components: a deterministic linear function
of time, a purely continuous random component (a multiple of Brownian
motion) and a purely discontinuous random component (a compound Poisson
process). In this case the coefficient of the Brownian component is zero and
the jumps of the compound Poisson process all have size 1.
(ii)
By the independent increments property, Nt
(iii)
(a)
E[Yt
| Ft ] e
[u c (t s )]
This is equal to Yt if c
(b)
E[e
Nt
Nt ~ Po( s ) given Nt.
| Ft ] e
[u c (t s ) N t ]
s[ e
1]
[1 e ] .
If c < , the line y = c increases slowly from 0 as increases,

whereas [1 e ] increases more rapidly to start with, but never
April 2004
Examiners Report
exceeds . Therefore there must be a crossing point at a positive value

of .
If, on the other hand, c > , then there is no positive crossing point; the
line y = c goes slowly to
as
, whereas [1 e ] tends to
exponentially fast as
.
(iv)
(c)
Yes; this is the condition for the company to remain solvent.
(a)
If S hits K first, then ST = K exactly, since increases in S occur

continuously. But if S becomes negative before hitting K, then the
negative movement must have been caused by a downward jump;
jumps are of magnitude 1, so in this case ST can be no less than 1.
(b)
Since St is between 1 and K for all 0
T, it follows that Ymin(t ,T ) is
bounded above and below, from which we deduce that the optional
stopping theorem applies.
(c)
(d)
ST = K with probability 1 , or with probability , ST takes a value

somewhere between 1 and 0. Therefore YT takes the value e K with
probability 1 , or with probability it takes some value between 1
and e . Hence the result.
Clearly,
E[YT ] e
1 e
1 e
As K becomes larger, the fact that

limit we have
e u.
.
< 0 means that e
0, so in the
There was an error in the final part of the question; where the question asked for a lower
bound, the quantity which could be derived from the previous part of the question was in fact
an upper bound. Candidates who reached the last part and were confused by the error were
treated generously.
It appeared that many candidates tried this question when they were short on time. Those
candidates who attempted part (iv) often did quite well on it, even if they had omitted earlier
parts of the question.
Page 13
EXAMINATIONS
Subject 103

1.
booklet.
2.
supervisor.
3.
4.

question paper.
103
S2004
A sequence x1, x2,

, xn of observations is provided, each of which takes one of a
finite set, S, of possible values. A researcher wishes to fit a discrete-time Markov
chain model to these data.
(i)
Write down estimates of the transition probabilities of the Markov chain.
[1]
(ii)
Outline the purpose of model validation.
[1]
(iii)
Describe the role of simulation in sensitivity analysis in the given context. [3]
[Total 5]
A person who catches a particular virus for the first time immediately falls ill with
Disease A. The illness lasts for a random length of time, whose mean depends on the
age of the person. Having recovered from Disease A, the person will not fall ill with it
again. However, if the person catches the same virus again after the age of 18, the
result will be Disease B, which will affect the person for an average of three months
regardless of age. After suffering from Disease B the person will not catch the virus
again.
It is proposed that a continuous-time Markov model can be used to model the person s
medical history, with states:
0: not currently ill with Disease A or Disease B
1: currently ill with Disease A
2: currently ill with Disease B
(i)
Explain why the proposed state space is inadequate and suggest an enlarged
state space which can support a Markov model.
[2]
(ii)
Draw a diagram to illustrate the possible transitions.
(iii)
State, with reasons, whether a time-homogeneous or a time-inhomogeneous

model is more suitable to model the medical history of a single person.
[1]
(iv)
Explain why a national medical service might find that a time-homogeneous

model was just as good as a time-inhomogeneous one for planning the
provision of treatment for Diseases A and B.
[1]
[Total 6]
103 S2004
[2]
You are given a stream of standard identically and independently distributed uniform
[0,1] random variables U1, U2, U3, ..
Describe how to use this random variable stream to generate random variates with the
following distributions:
(i)
The discrete distribution with possible values A, B and C with respective

probabilities 1/5, 1/4 and 11/20.
(ii)
[2]
The continuous distribution with probability density function

f ( x) =
(iii)
(10 x) /18 for 4

0
x 10
[3]
otherwise
The continuous distribution function with probability density function
2
f ( x) =
sin 2 x for 0
0
[3]
otherwise
[Hint: Use the Acceptance-Rejection method.]

[Total 8]
(i)
Define a standard Brownian motion {Bt t
(ii)
Assume that St , the price of a share at time t , satisfies the stochastic

differential equation
dSt = St dt
St dBt , t
0} .
0.
Solve the equation for St in terms of S0 and Bt .
[3]
For the values = 20% and = 15% on an annual basis, calculate the
probability that the price of the share will exceed 120 in three months time if
the current price of the share is 96.
[4]
[Total 9]
(iii)
103 S2004
[2]
PLEASE TURN OVER
A bag contains N balls, all green or red. At each stage a ball is taken out of the bag at
random and is replaced by a ball of the other colour. Let Xn denote the number of
green balls in the bag after n stages.
(i)
Explain why Xn, n 0 is a discrete time Markov chain with state space
S = {0, 1, 2, .., N} and with transition probabilities
pi ,i
N i
and pi ,i
N
i
N
[2]
(ii)
Show that the Markov chain Xn, n

period.
(iii)
Show that for this process the stationary distribution

is given by
(iv)
2N
0 is irreducible and periodic and state its

[3]
for i = 0, 1, 2,
= ( 0,
1,
., N.
2,
..,
N)
[4]
State, with a reason, whether it is the case that

lim P( X n = j | X 0 = 0) =
j.
[1]
[Total 10]
A model frequently used for interest rates is the Cox-Ingersoll-Ross process, which is
an It process Xt satisfying the stochastic differential equation
dX t = (b X t )dt
X t dBt ,
where Bt is a standard Brownian motion.

(i)
Define m1(t) = E(Xt | X0 = x).

(a)
Verify that m1 satisfies the ordinary differential equation
dm1
= (b m1 )
dt
(b)
103 S2004
Solve the equation to determine m1(t) for all t
0.
[3]
Define Yt = X t2 and m2 (t ) = E[ X t2 | X 0 = x] .
(ii)
(a)
Use It s Lemma to derive a stochastic differential equation satisfied

by Yt.
(b)
Deduce an ordinary differential equation satisfied by m2(t).
(c)
Assuming that
limt
dm2
dt
0 as t
Var[ X t | X 0 = x] =
b
2
, show that
2
.
[7]
[Total 10]
A stationary autoregressive process X t is defined by the recursive relationship

Xt =
1( X t 1
2(Xt 2
p ( Xt p
) et
where {et t 1} is a sequence of independent, zero-mean Normal variables, each

with variance
(i)
Derive the Yule-Walker equations

k
for 0 k
2
1 k 1
2 k 2
p , where
p k p
= Cov( X t X t
1k
k).
[2]
(ii)
Describe a diagnostic procedure based on a sequence of observations from a

time series for testing whether the underlying time series can be modelled as a
second order autoregressive process.
[2]
(iii)
Consider the second order autoregressive process

Xt = 0 6Xt
103 S2004
0 3Xt
et
(a)
Determine whether the process can be stationary.
(b)
State, with a reason, whether the process possesses the Markov

property.
(c)
Assuming that
= 1 , calculate the values of
2.
[7]
[Total 11]
PLEASE TURN OVER
(i)
Write down the defining equation of an ARMA(1,1) process, identifying the

parameters of the process.
[2]
(ii)
Explain what it means to say that a time series is stationary and state (but do
not prove) a condition needed to ensure that an ARMA(1,1) process can be
stationary.
[2]
(iii)
Outline the method of moments parameter estimation technique as it would be

applied to estimate the parameters of an ARMA(1,1) process.
[3]
(iv)
Suppose an individual has fitted the following model to a dataset

xt = 9 12 0 71xt
et
0 17et
The most recently observed value in the series is x25 = 14 82 , with estimated
residual e 25
1 98 .
(a)
Obtain estimates x 25(1) and x 25(2) for x26 and x27 .
(b)
The simplest form of exponential smoothing used at time 24 gave a

forecast for x25 of 12.97. Assuming the smoothing parameter is equal
to 0.3, find the forecast for x26 .
[4]
(v)
Discuss when the method of exponential smoothing might in practice be

preferred to a method based on the Box-Jenkins technique.
[2]
[Total 13]
Vehicles in a certain country are required to be assessed every year for roadworthiness. At one vehicle assessment centre, drivers wait for an average of 15
minutes before the road-worthiness assessment of their vehicle commences. The
assessment takes on average 20 minutes to complete. Following the assessment, 80%
of vehicles are passed as road-worthy allowing the driver to drive home. A further
15% of vehicles are categorised as a minor fail ; these vehicles require on average
30 minutes of repair work before the driver is allowed to drive home. The remaining
5% of vehicles are categorised as a significant fail ; these vehicles require on
average three hours of repair work before the driver can go home.
A continuous-time Markov model is to be used to model the operation of the vehicle
assessment centre, with states W (waiting for assessment), A (assessment taking
place), M (minor repair taking place), S (significant repair taking place) and H
(travelling home).
(i)
(ii)
103 S2004
Explain what assumption must be made about the distribution of the time
spent in each state.
[1]
Write down the generator matrix for this process.
[2]
(a)
Use Kolmogorov s Forward Equations to write down differential

equations satisfied by pWM(t) and by pWA(t).
(b)
Verify that pWA (t ) = 4e

minutes.
(c)
Derive an expression for pWM(t) for t
t / 20
4e
t /15
for t
0, where t is measured in
0.
[7]
(iii)
Let Ti be the expected length of time (in minutes) until the vehicle can be
driven home given that the assessment process is currently in state i.
(a)
Explain why TW = 15 + TA.
(b)
Derive corresponding equations for TA, TM and TS.
(c)
Calculate TW.
[4]
[Total 14]
10
Assume that the spot rate of interest at time t , S (t ) , can be modelled by

S (t ) = e 2 W (t ) , where W (t ) is a Wiener process with drift coefficient
diffusion coefficient 1 such that W(0) = 0.
and
(i)
Write down an expression for W (t ) in terms of a standard Brownian motion

B (t ) .
[1]
(ii)
Show that {S (t ) t
(iii)
Let Ta = inf{t S (t ) = a} for some 0 a 1 .
0} is a continuous-time martingale.
1 as t
[4]
(a)
Prove that P[ S (t ) a]
(b)
Deduce that P[Ta
(c)
Explain why the fact that E[ S (Ta )] S (0) does not contradict the
optional stopping theorem.
] = 1 and that E[ S (Ta )] = a .
[5]
(iv)
Now suppose instead that a 1 and define Ta as before.

(a)
Explain why E[ S (Ta )] = aP[Ta
].
(b)
Apply the optional stopping theorem to find P Ta
.
[4]
[Total 14]
END OF PAPER
103 S2004
EXAMINATIONS
September 2004
Subject 103
EXAMINERS REPORT
September 2004
Examiners Report
he examiners were pleased to note that the overall quality of answers on this
final sitting of subject 103 was high and that many of the candidates
demonstrated a good knowledge of the principles and practice of stochastic
modelling. As always, credit was awarded for comments which showed that
candidates had an understanding of the topic covered in a question, even if the
calculations gave the wrong answer due to some mathematical error.
Question 7 was particularly well answered, with Questions 2 and 4 not far
behind. Questions 6 and 10 had the lowest proportion of good answers; it is
possible that time pressure played a role in the case of Question 10.
(i)
Let nij denote the number of direct transitions from state i to state j, with ni+
the total number of transitions out of state i. Then pij nij / ni .
(ii)
Model fitting aims to find the best-fitting model in a given class. But it is
conceivable that even the best-fitting model in the class does not fit very well.
Model validation is a set of procedures to test the adequacy of the fit.
(iii)
Sensitivity analysis is part of model validation. The purpose is to determine

whether the behaviour of the fitted model would be substantially different if
the parameter values were slightly different from the estimates already
obtained.
The technique involves simulating the fitted process a large number of times,
using several simulations for each of a number of slightly different parameter
values, then examining the output of the simulation to attempt to identify
systematic differences.
It is important that the same sequence of random numbers be used in each of
the sets of simulations to ensure comparability.
Many candidates failed to mention the importance of using the same sequence
of pseudo-random numbers. Apart from that, most answers showed good
knowledge of the principles of modelling.
(i)
It is inadequate because someone who has never suffered from disease A or B

is not in the same position as someone who has suffered from one or both in
the past but is currently healthy.
The state space should be extended by splitting state 0 into 3: 0: Has never
suffered from A or B , A: Has suffered from A but is now healthy ,
AB: Has suffered both A and B but is now healthy.
(ii)
0
1
Page 2
AB
A
2
September 2004
Examiners Report
(iii)
Only a time-inhomogeneous model can properly reflect the dependence of

both the recovery rate for A and the infection rate for B on the age of the
person.
(iv)
If, in a population taken as a whole, the number of people in each age group is
roughly constant over time, then the age-dependent transition rates of the
individuals who make up the population can be averaged out to give a timehomogeneous model which works perfectly adequately given that national
medical services are generally only concerned with total numbers falling ill.
This question was answered well in general. Where candidates lost marks it
was often due to mis-specifying the additional states in part (i). Splitting state
A into Has recovered from Disease A and is aged below 18 and Has
recovered from Disease A and is aged 18 or more is reasonable when
modelling an entire population, but does not lead to a time-homogeneous
Markov model when applied to a single individual, since one s 18th birthday
does not occur at a random time. However, answers along these lines with
good explanations were given full marks.
(i)
Set Xi as follows:
A if 0 U i
Xi
(ii)
1/ 5
B if 1/ 5 U i 9 / 20
C if 9 / 20 U i
Use the inverse transformation method.
The distribution function is F ( x)
x 10
4
18
t dt
10 x
36
for 4
x 10 .
Solving the equation gives F 1 (u ) 10 6 1 u

So we can set X i
10 6 1 U i
or alternatively we could use X i

(iii)
10 6 U i
Use acceptance-rejection method:

Let V1 = U1 , so that V1 is uniformly distributed on [0, ] and has density
function g(x) = 1/ over that range.
We define
Page 3
September 2004
sup 2sin 2 x
sup f ( x) / g ( x)
0 x
Examiners Report
2.
0 x
If U2 < sin2 V1 let X1 = V1; otherwise reject this value and select a new pair
U1, U2. Repeat for other Xi
Answers to parts (i) and (ii) were generally good. For part (iii) many
candidates only described the general theory without specifying g(x) or
calculating the constant C.
(i)
A standard Brownian motion {Bt } is defined by the following properties:
B(0) 0 and Bt has independent increments; Bt Bs is independent of

Bu for 0 u s and s t .
Bt has stationary and Gaussian increments; Bt Bs N (0 t s ) .
Bt has continuous sample paths, i.e. t Bt is continuous.
(ii)
Using It s Lemma gives
d (log St )
1
dSt
St
1 1
(dSt ) 2
2
2 St
2
dt
dBt
dt
This implies that

2
log St
log S0
or, finally,
St
(iii)
Page 4
We have
S0 e
2
2
Bt
Bt
September 2004
Examiners Report
P St
x S0
Bt
P Bt
2
1
log
1
x
log
y
2
2
x
y
x
y
log
Substituting the values x 120 y

above we find that the answer is
96
02
0 15 and t
0 25 years
(2.3461) 1 0 9905 0 0095.
The calculations in parts (ii) and (iii) were well done. The definition in part
(i) caused some problems: it is necessary to mention the stationary,
independent increments property; then either of the two remaining properties
(continuous sample paths, normally distributed increments) implies the other.
(i)
The Markov property is clear: the chain jumps either up or down by 1, with
probabilities depending only on the current state, not the past history.
P( X n
i 1| X n
i ) is the probability that the (n+1)th ball selected is red,
which is just 1/N of the number of red balls at time n, which is N

(ii)
i.
From any state i it is possible to reach any other state j in just |j i| steps,
either all upwards or all downwards. This means that the chain is irreducible.
Every transition takes the chain from an even state to an odd one or vice versa,
which implies that the period must be an even number.
On the other hand, starting from state 0 it is possible to return to 0 in two
steps. Therefore state 0 has period 2 and, by irreducibility, all states have
period 2.
(iii)
To find the stationary distribution we can use the relationship suggested by the
Detailed Balance Equations:
i
N i
N
i 1
for i = 0, 1, 2,
N
i 1
., N
1.
Thus we get the recursive relationship

i 1
N i
i 1
for i = 0, 1, 2,
., N
Page 5
September 2004
Examiners Report
Starting with i = 0 and working forwards we get

1
N
1
N 1
2
N 2
3
N ( N 1)
(2)(1)
N ( N 1)( N 2)
(3)(2)(1)
and in general
N ( N 1)( N 2)
i (i 1)(i 2)
( N i 1)
(2)(1)
N!
( N i )! i !
N
i
Alternatively, write down the transition matrix P and use the equation
T
to obtain
1
N
2
N
N 1
N
To find
3
N
0,
i.e.
N ( N 1)
2
N!
1 ( N i )! i !
(iv)
2N
P=
N ( N 1)( N 2)
3!
+
N
i 0
Therefore the stationary distribution
we use the fact that

N
for i = 0, 1, 2,
N!
i!
, etc.
..+
= ( 0,
2N
1,
2,
= 1
1
2N
..,
N)
is given by
., N
P ( X n j | X 0 0) does not converge, being alternately zero and non-zero,

since X is periodic.
The derivation of the stationary distribution in part (iii) caused difficulties
with many candidates, but otherwise candidates showed a good understanding
of discrete-time Markov chains.
Page 6
(i)
Examiners Report
dm1 d
1
Ex X t
Ex [dX t ]
Ex [b X t ]
(b m1 ) , where Ex
dt
dt
dt
denotes conditional expectation given X0 = x. This derivation uses the
fact that the increments of Brownian motion have expectation equal to
zero.
(a)
d
[e t m1 (t )]
be t , implying that
dt
m1 (t ) e t [ x b(e t 1)] b ( x b)e
(b)
(ii)
September 2004
dYt
(a)
(dX t ) 2
2 X t dX t
[2 b
2 X t [ (b X t )dt
X t dBt ]
X t dt
] X t dt 2 X t2 dt 2 X t3/ 2 dBt
d
2
m2 (t ) [2 b
]m1 (t ) 2 m2 (t ). Again we have used the fact
dt
that Brownian increments have mean zero.
(b)
(c)
We do not need to solve the equation, but just to note that since dm2/dt
tends to 0, this implies that 2 limt
m2(t) = [2 b + 2] limt
m1(t)
= 2 b2 + b 2. Therefore
b 2
limt
E[ X t2 | X 0 x] b 2
, from which we deduce that
2
b 2
.
limt
Var[ X t | X 0 x]
2
This question was relatively poorly answered, although much of it is based on

the standard theory of Ordinary Differential Equations.
(i)
Taking covariances with X t
Cov( X t X t
k)
1Cov( X t 1
Xt
k)
for k 1 in (1) gives

2Cov ( X t 2
Xt
k)
p Cov ( X t p
Xt
k)
which gives the Yule-Walker equations since, by definition,

Cov( X t X t k ) for 0 k p .
k
For k
(ii)
0 , there is an extra term which accounts for Cov( X t et )
A diagnostic test is based on the partial ACF and uses the fact that, for an
AR(2) process, the partial autocorrelations, k , are zero for k 2 .
Page 7
The values of
September 2004
are estimated by the partial ACF,
asymptotic variance of
(a)
, and for k
2 the
is 1 n . Using a normal approximation, values of the
sample partial ACF outside the range 2

may be inadequate.
(iii)
Examiners Report
n indicate that the AR(2) model
The process can be written in terms of the backward shift operator as

(1 0 6 B 0 3B 2 ) X t et .
Hence the characteristic polynomial is 1 0 6 z 0 3 z 2 with roots
06
(0 6) 2 1 2
i.e. the roots are 1
06
156 6 .
Since both roots lie outside the unit circle, the process can be
stationary.
(b)
X t is not Markov since the conditional distribution of X k

history up to time k depends on X k 1 as well as on X k .
(c)
The Yule-Walker equations in this case yield

0
06
03
(3)
06
03
(4)
06
03
(5)
From (4) we have
07
06
6 0
7
(6)
and substituting into (5) we get

2
36
70
3
10
57
70
(7)
Inserting the last two equations into (3) we obtain

0
36
70
171
700
which gives
Page 8
36
70
171
700
700
169
given the
September 2004
Examiners Report
Then (6) and (7) yield resp.

1
6 0
7
600
169
570
169
The examiners were pleased to note the high quality of answers to this
question. It appears that the theoretical principles of Time Series analysis are
well understood.
(i)
The equation is
Xt
( Xt
) et
et
The parameters are (the autoregressive parameter), (the moving average

parameter), the mean level and the innovation standard deviation .
(ii)
A time series process is (weakly) stationary if the mean of the process,

mt E ( X t ) , does not vary with time and the covariance of the process,
Cov( X t X s ) depends only on the time difference t s and not on the
particular values t s .
1 is needed.
For the model in (1) to be stationary,

(iii)
For the method of moments, we calculate the theoretical ACF

of the parameters
in terms
. Then we find the sample ACF, say r1 r2 from the data.
Subsequently we obtain estimates for
by equating
with r1 and
with r2 .
The value of 2 is estimated using the calculated value of
variance, whereas an estimate for is the sample mean x .
(iv)
(a)
and the sample
Using the given values we obtain the forecasts
x 25(1) 9 12 0 71(14 82) 0 17( 1 98) 19 306

and
x 25(2) 9 12 0 71(19 306)

(b)
22 827
For exponential smoothing the equation is

x 25(1)
x 24(1)
x25
x 24(1)
12 97 0 3(14 82 12 97) 13 525
Page 9
(v)
September 2004
Examiners Report
Exponential smoothing is simple to apply and does not suffer from problems
of over-fitting. If the data appear fairly stationary but are not especially well
fitted by any of the Box-Jenkins methods, exponential smoothing is likely to
produce more reliable results. More advanced versions of exponential
smoothing can cope with varying trends and multiplicative variation.
Many candidates omitted to mention as a parameter in part (i). Marks for
this question were not quite as good as for Q7, indicating that the practical
aspects of Time Series analysis are less well understood than the theoretical
ones.
(i)
The Markov model implies that holding times are exponentially distributed.
(ii)
The generator matrix is as follows (in minutes then, equivalently, in hours):

W
1
400
1
25
1
30
1
180
3
400
1
30
1
180
W
A
M
S
H
(a)
1
15
1
20
1
15
4
0
0
0
4
0
0
0
3 0.45 0.15 2.4
0
2
0
2
1
1
0
0
3
3
The equations are as follows (first if t is in minutes, then in hours)
d
pWM (t )
dt
d
pWA (t )
dt
d
pWM (t )
dt
d
pWA (t )
dt
Page 10
1
3
pWM (t )
pWA (t )
30
400
1
1
pWA (t )
pWW (t )
20
15
2 pWM (t ) 0.45 pWA (t )
3 pWA (t ) 4 pWW (t )
(b)
September 2004
Examiners Report
First note that pWW(t) = e t/15. Then try inserting the given formula in
the second equation above:
LHS
d
pWA (t )
dt
1
e
5
t / 20
4
e
15
t /15
and
RHS
1
1
pWA (t )
pWW (t )
20
15
1
e
5
1
e
5
t / 20
t /15
1
e
15
t /15
which is equal to the RHS, as required.

We should also check that the formula gives pWA(0) = 0, which it does.
(c)
d t / 30
3 t / 30
e
pWM (t )
e .4 e t / 20 e t /15 with pWM(0) = 0
dt
400
t / 30
implies that e
pWM (t ) 0.9 1.8e t / 60 0.9e t / 30 , which simplifies
to pWM (t ) 0.9e
(iii)
t / 30
1.8e
t / 20
0.9e
t /15
(a)
The expected length of time spent in state W is 15 mins, after which

there is a transition to state A with probability 1.
(b)
The other equations are:
(c)
TA
= 20 + 0.15 TM + 0.05 TS
TM
= 30
TS
= 180
Solving these equations gives TA = 20 + 4.5 + 9 = 33.5 and TW = 48.5

mins
A number of students suggested that pWW (t ) 1 pWA (t ) , which works in the

two-state case. In this example, however, it is only possible to state that
pWW (t ) pWA (t ) pWM (t ) pWS (t ) pWH (t ) 1, which is not the same.
It was disappointing that not many candidates made substantial progress with
solving the differential equations, but the last part of the question was in
general well done.
10
(i)
The Wiener process can be defined as Wt
(ii)
We need to show that
Bt . In this case
1.
Page 11
E St S s
Ss
September 2004
s t
as well as proving that E[ St ]

But W (t )
Examiners Report
N ( t t ) , so that M t ( x) e
x tx 2 2
We have
E St S s
E e
e
E e
2 W (t )
Ss
2 W (t ) W ( s ) W ( s )
2 W (s)
E e
Ss
2 (W (t s ))
(1)
using stationarity and independence of the increments.

From the definition of a Wiener process with drift above, we have
W (t s )
N ( (t s ) (t s ))
(2)
so
E e
where M t
2 (W (t s ))
Mt s ( 2 )
is the moment generating function of the normal distribution in
(2). But for this distribution we know that M t s ( x) e
(t s ) x (t s ) x 2 2
Therefore
Mt s ( 2 ) e
(t s ) (t s )(2 )2 2
now (1) shows that E St S s
S s as required.
Finally, check the expectation: St

E[ S t ]
(iii)
Page 12
(a)
E[ St ] E[e
2 W (t )
0 , so
] Mt ( 2 ) e
t ( 2 ) t ( 2 )2 2
P[ S (t ) a ] P[e 2 ( B (t ) t ) a ] P[ 2 B(t ) b 2 2t ] , where

b = log a. If
0 , this becomes
September 2004
b
2
P B(t )
0 we have
whereas in the case

b
2
P B (t )
Examiners Report
b
2
and in both cases the RHS tends to 1 as t

(b)
P[Ta
t ] P[ S (t )
a]
By definition, S (Ta )
(iv)
0 as t
a . Therefore E[ S (Ta )] a .
(c)
This is not a contradiction because the conditions of the optional

stopping theorem are not satisfied. Neither S (t ) nor Ta is bounded
above, even though S (t ) is a convergent martingale.
(a)
In this case Ta is only finite if S (t ) hits a , which is not certain.

However, as above it is certain that S (t ) 0 almost surely.
Therefore
S (Ta )
a
0
if Ta
if Ta
It follows that E[ S (Ta )] aP[Ta

(b)
].
Now the optional stopping theorem applies, since S (t Ta ) is bounded

below by 0 and above by a .
We may deduce that
1
a
In part (ii) a large number of candidates did not even attempt to prove that
E (| St |)
: this condition is a requirement for S to be a martingale and
should not be omitted. However, most candidates had a good idea of how to
prove that S satisfied the conditional expectation condition.
Parts (iii) and (iv) attracted at best sketchy answers. The examiners were
unsure whether this was due to pressure of time or to lack of familiarity with
applications of the optional stopping theorem.
1 S (0)
aP[Ta
i e that P[Ta
END OF EXAMINERS REPORT
Page 13

Stochastic Modelling 2000-2004

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Stochastic Modelling 2000-2004

Uploaded by

Copyright:

Available Formats

Faculty of Actuaries

Time allowed: Three hours

Mark allocations are shown in brackets.

Attempt all 11 questions, beginning your answer to each question on a

Hand in BOTH your answer booklet and this question paper.

Let Xn = Y1 + Y2 + ... + Yn be a simple random walk with step distribution

Let X and Y be jointly normal random variables with means X , Y , variances

The total number of claims received by an insurance company is described by

According to a model used in econometrics, three economic time series, X, Y

whether X, Y and Z are I(0) or I(1)

whether each of X, Y and Z individually satisfies the Markov property

whether the vector-valued process {(Xn , Yn , Zn) : n 1} satisfies the

whether X and Y are cointegrated

Let Bt be a standard Brownian motion, and let Ft = (Bs , 0 s t) be its natural

Derive the conditional expectations E[ Bt2Fs ] and E[ Bt4Fs ] , where s t.

Hence construct a martingale out of Bt4 .

Apply the inverse transform method to generate an observation from

using a pseudo-random number u in the range 0 < u < 1. Explain how

The Cauchy distribution has density function

where is a positive parameter.

PLEASE TURN OVER

Calculate the autocovariance function {k : k 0} and

State the conditions on the values of the parameters such that

A sequence of observations x 1 , x 2 , ..., x n has sample variance $ 0 = 14.5,

The members of a health insurance scheme are classified as contributors or

Construct a discrete time Markov chain to model this health

Draw the transition graph.

Write down the transition matrix of the chain.

Explain whether the above Markov chain is irreducible, periodic or

Calculate the stationary probability distribution of the chain.

Determine the proportion of beneficiaries among the

Explain how b, c and f should be related if the scheme is to be

Calculate the average profit per period per member in the

A stationary second-order autoregressive process X, which may be assumed to

Obtain an equation for 1 in terms of 0 and 2 by substituting for

Derive similar equations for 2 and 0.

State the autocorrelation function k of X for k = 0, 1, 2.

Define standard Brownian motion Bt , t 0 and give its transition

Write down the transition probability density of general

Let St defined represent a share price at time t.

Solve the stochastic differential equation

Calculate, given the parameters = 25% p.a., = 20% on an annual

where y 0 and G denotes the normalised Gaussian probability

PLEASE TURN OVER

Write down the matrix of transition rates, {ij : i, j = A, F, I, O, D}, of

Calculate the proportion of patients who eventually receive in-patient

Derive expressions for the probability that a patient arriving at time

yet to be classified by the junior doctor at time t, and

undergoing further investigation at time t

Explain in words why mi satisfies the following equation:

Hence calculate the expectation of the total time until discharge

Describe briefly what additional feature of the data might be used to

The hospital management committee believes that replacing the junior

Outline briefly the calculations that would need to be performed

Discuss whether the current model is suitable as a basis for

Subject 103 (Stochastic Modelling) April 2000 Examiners Report

Assume E[XY] = a + bY and determine a, b using:

orthogonality condition E{(X - E[XY])Y} = 0.

(i) gives E[XY] - aE[Y] -bE[Y2] = 0;

we have E[XY] = rsX sY + mX mY and (i) yields

(ii) gives a + bmY = mX .

Subject 103 (Stochastic Modelling) April 2000 Examiners Report

P(s, t) = P(s, t) A(t),