Faculty of Actuaries Institute of Actuaries

EXAMINATIONS

September 2000

Subject 103 — Stochastic Modelling

EXAMINERS’ REPORT

ã Faculty of Actuaries
ã Institute of Actuaries
 Subject 103 (Stochastic Modelling) — September 2000 — Examiners’ Report

1 (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)
σm,m+1 = λp .
P

(ii) The expected total infection time is the sum of the reciprocal rates
P
λp
åN −1
m =1
1
m( N − m)
.

1 1 é1 1 ù
(Since = ê + ú , this time may also be expressed as
m( N − m) N ëm N − m û
N −1
λp m å
N −1 1
=1 m
.)

2 (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+ = Σj 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 Xn−1. 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.

3 (i) Since {Un} is Gaussian and stationary, it is determined uniquely by its

mean and autocovariance functions. For k > 0 we have γk = Cov(Un , Un−k)
τ2 −θk τ2
= e , so that the ACF is ρk = e−θk and the variance γ0 = .
2θ 2θ

σ2
(ii) Compare this with the corresponding values for an AR(1): γ0 = and
1 − α2
ρk = αk for k > 0.

τ2
The two are seen to match as long as α = e−θ and σ2 = (1 − e−2θ) .
2θ

Page 2
 Subject 103 (Stochastic Modelling) — September 2000 — Examiners’ Report

4 (i) If X satisfies dXt = Yt dBt + Zt dt, then f(Xt) satisfies

d[f(Xt)] = f ′(Xt) Yt dBt + { f ′(Xt) Zt + ½ f ′′(Xt) Yt2 } dt.

(ii) (x4)′ = 4x3 , (x4)′′ = 12x2.

d( Bt4 ) = 4 Bt3 dBt + 1

2
. 12Bt2dt = 4 Bt3 dBt + 6 Bt2dt.

(iii) t
ò0 d( Bs4 ) = 4 ò t0 Bs3 dBs + 6 òt0 Bs2 ds.

Therefore t
ò0 Bs3 dBs = 1
4
Bt4 − 3 t
ò
2 0
Bs2 ds.

æ1 2 1 ö æ8 3 5ö æ 29 21 14 ö
1ç ÷ 1 ç ÷ 1 ç ÷
5 (i) P = ç2 0 2÷ , P = 2
8 6 2÷ , 3
P = 26 18 20 ÷ .
4ç ÷ 16 çç ÷ 64 çç ÷
è3 1 0ø è5 6 5ø è 32 15 17 ø

14 7
(ii) (a) P133 = = = 0.21875.
64 32

2 2 1
(b) P23 = = = 0.125.
16 8

14 3 9 3 8 3 14 × 14 + 9 × 20 + 8 × 17 512 8
(c) P13 + P23 + P33 = = = .
31 31 31 31 × 64 31 × 64 31

(iii) By time n = 300 the effects of the starting point have worn off. The
answer is therefore indistinguishable from the stationary probability π3 in
all three cases.

It is easily observed that the distribution in (c) is stationary, so that

8
π3 = .
31

6 (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
 Subject 103 (Stochastic Modelling) — September 2000 — Examiners’ Report

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

sn − sn−1 should be independently normally distributed with
constant mean and variance:

a test of normality (Anderson-Darling, Kolmogorov-Smirnov, χ2

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) (a) A Lévy process is the sum of three independent components: a

deterministic part of the form µ + αt, a Brownian part of the form
σBt 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.

7 (i) (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, σ2e ) is the standard choice: one method of achieving this is to
define, for each integer t,

e2t = σe −2 log u2t sin(2πu2t +1 )

e2t+1 = σe −2log u2t cos(2πu2t +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) 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.

(ii) The models do not possess the correct correlation structure.

Page 4
 Subject 103 (Stochastic Modelling) — September 2000 — Examiners’ Report

(iii) (a) ρ1 = Corr(Xt ,Xt−1) = α1Corr(Xt−1 ,Xt−1) + α2Corr (Xt−2 ,Xt−1) = α1 + α2 ρ1 .

Hence ρ1 = α1 / (1 − α2)

ρ2 = Corr(Xt ,Xt−2) = α1Corr(Xt−1 ,Xt−2) + α2Corr(Xt−2 ,Xt−2) = α1 ρ1 + α2 .

(b) We have 0.7 = ρ1 = α1 / (1 − α2)

and 0.5 = ρ2 = α2 + α12 / (1 − α2) = α2 + 0.7α1. Two equations in two

35 1
unknowns. Solution: α1 = , α2 = .
51 51

(2 marks for the observation that α1 = 0.7 and α2 = 0 is very close

to giving the right answer, as it gives ρ2 = 0.49.)

8 (i) ′ (t ) = µP0,1(t) − λP0,0(t), or a more general form such as P0,0

P0,0 ′ (t ) =
ΣP0,k(t)σk,0

(ii) Since P0,1(t) = 1 − P0,0(t),

′ (t ) = µ(1 − P0,0(t)) − λP0,0(t). Any solution method will do,

we have P0,0
d (λ+µ)t µ
e.g. [e P0,0(t)] = µe(λ+µ)t , solved by P0,0(t) = + Ce−(λ+µ)t, with C
dt λ+µ
being determined by the fact that P0,0(0) = 1.

t t
(iii) E0Ot = E0 ò t0 Is ds = ò0 E0 Is ds = ò0 P0,0(s)ds

µ λ
= t+ (1 − e−(λ+µ)t)
λ+µ (λ + µ) 2

(iv) Since the process must be in state 0 or state 1 at all times, the solution is
λ λ
just t − E0 0t = t− (1 − e−(λ+µ)t).
λ+µ ( λ + µ )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

æ λ λ ö
γt + β ç t− (1 − e −( λ +µ )t ) ÷
èλ +µ (λ + µ)2
ø

æ µ λ ö
and αç t+ (1 − e −( λ+µ )t ) ÷ .
èλ +µ (λ + µ)2
ø

Page 5
 Subject 103 (Stochastic Modelling) — September 2000 — Examiners’ Report

In the long run, then, as t → ∞, we require αµ = βλ + γ(λ + µ) to

break even.

(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.

9 (i) If Vt is a martingale, then its expectation must be constant and equal to

its initial value euy.

Therefore Eeu( y+µt +αBt )−c(u )t = euy +(uµ+u α

2 2
/ 2− c ( u ))t
= euy.

Thus we must have c(u) = uµ + u2α2 / 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 Mt∧T 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{α } 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
 Subject 103 (Stochastic Modelling) — September 2000 — Examiners’ Report

t
PAA (s, t) = e − µxdx 2
− s2 ) / 2
10 (i) òs
= e −µ(t

2
− s2 ) / 2 2
(ii) P[Rs ≥ w] = PAA (s, s + w) = e −µ(( s +w ) = e −µsw−µw /2
. Therefore

2
∞
E[Rs] = ò0 e −µsw−µw /2
dw.

Complete the square at the exponent to get

2 2 2 2 dx 1 1 − G( s µ )
E[Rs] = e µs /2
ò0
∞
e −µ( s+w ) /2
dw = e µs /2
ò
∞
s µ
e −x /2
= .
µ µ g( s µ)

(iii) From (ii) and the given bound

1 1 1
E[ Rs ] ≤ =
µ s µ sµ

1 æ 1 1 ö 1 1
E[ Rs ] ≥ ç − ÷ = − 3 2.
µ çs µ
3 3/2 ÷
sµ ø sµ s µ
è

The first inequality yields

1 1
µ≤ = = 0.00238 . (year)−2.
sE[ Rs ] 420

The second inequality can be written as

µ 1
µ2E[Rs] − + ≥ 0,
s s3

so µ must lie outside the interval:

1 1 4 E[ Rs ] 4E[ Rs ] 24
± 2
− 3
1± 1− 1± 1−
s s s s 70
= = = [0.00023, 0.00215]
2E[ Rs ] 2sE[ Rs ] 2 × 6 × 70

1
In fact, since clearly µ − we see that µ must lie in the interval
sE[ Rs ]
[0.00215, 0.00238].

(iv) Use the inverse transform method, X = F−1(U).

In this case 1 − F(x) = exp( − ò 70

x
[ a + bt ] dt ) = exp{−a(x − 70) − ½b(x2 − 702)}.

Page 7
 Subject 103 (Stochastic Modelling) — September 2000 — Examiners’ Report

Therefore ½bx2 + ax = −log(1 − F(x)) + ½702b + 70a. Replace F(x) by u to

get

−a + a 2 + 2b[ − log(1 − u) + ½702 b + 70a ]

x = F−1(u) =
b

= −r + 702 + 140r + r 2 − 2b−1 log(1 − u ),

where r = ab−1. 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 = α(xn − xn−1) + en+1 + βen .

(b) The parameters are α, β and σ2e . The trend removal process would
have accounted for any µ parameter.

(iii) xˆ n (1) = E(Xn+1xn , ..., x1) = xn + α(xn − xn−1) + E(en+1 + βenxn , ..., 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 − xn−1 − α(xn−1 − xn−2) − βen−1 , which may be iterated back to get
en in terms of the known x and the known e0.

Thus

xˆ n (1) = xn + α(xn − xn−1) + βen .

Similarly,

xˆ n (2) = E(Xn+2Fn) = E(Xn+1 + α(Xn+1 − xn) + en+2 + βen+1Fn )

= (1 + α) xˆ n (1) − αxn .

We see that Xn+1 − xˆ n (1) = en+1 , so that the prediction variance is just
Var(en+1) = σ2e .

Page 8
Subject 103 (Stochastic Modelling) — September 2000 — Examiners’ Report

(iv) Since en = xn − xˆ n −1 (1) , we have

xˆ n (1) = xn + α(xn − xn−1) + β(xn − xˆ n −1 (1));

if we set α = 0 and β = −ξ, 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