STAT 520

FORECASTING AND TIME SERIES

2013 FALL
Homework 02

1.
2.7.(a) Since {Yt } is a stationary process, it has constant mean µ over
time and autocovariance function γk which is free of t. Letting Wt = ∇Yt =
Yt − Yt−1 , we obtain the mean function of {Wt }

E(Wt ) = E(Yt − Yt−1 ) = E(Yt ) − E(Yt−1 ) = µ − µ = 0

which is constant over time t. The autocovariance function of {Wt }

Cov(Wt , Wt−k ) = Cov(Yt − Yt−1 , Yt−k − Yt−k−1 )

= Cov(Yt , Yt−k ) + Cov(Yt , −Yt−k−1 ) + Cov(−Yt−1 , Yt−k ) + Cov(−Yt−1 , −Yt−k−1 )
= Cov(Yt , Yt−k ) − Cov(Yt , Yt−k−1 ) − Cov(Yt−1 , Yt−k ) + Cov(Yt−1 , Yt−k−1 )
= γk − γk+1 − γk−1 + γk = −γk+1 + 2γk − γk−1

only depends on time lag k. So we conclude that {Wt } is stationary.

2.7.(b) From previous result, we know that the first difference of a sta-
tionary process is still stationary. In this case, since we proved that {Wt } is
stationary, then the first differences of {Wt }, {Ut }, are stationary as well.

2.9.(a) Suppose Yt = β0 +β1 t+Xt , where {Xt } is a zero-mean stationary

series with autocovariance function γk and β0 and β1 are constant. The mean
function of {Yt }

E(Yt ) = E(β0 + β1 t + Xt ) = β0 + β1 t + E(Xt ) = β0 + β1 t + 0 = β0 + β1 t

which is not constant over time t. Then {Yt } is not stationary.

Consider Wt = Yt − Yt−1 , we obtain the mean function of {Wt }

E(Wt ) = E(Yt − Yt−1 ) = E(Yt ) − E(Yt−1 )

= E(β0 + β1 t + Xt ) − E(β0 + β1 (t − 1) + Xt−1 )
= β0 + β1 t + E(Xt ) − β0 + β1 (t − 1) + E(Xt−1 )
= β0 + β1 t + 0 − β0 + β1 (t − 1) + 0 = β1

which is constant over time t.

1
 The autocovariance function of {Wt }
Cov(Wt , Wt−k ) = Cov(Yt − Yt−1 , Yt−k − Yt−k−1 )
= Cov{β0 + β1 t + Xt − [β0 + β1 (t − 1) + Xt−1 ],
β0 + β1 (t − k) + Xt−k − [β0 + β1 (t − k − 1) + Xt−k−1 ]}
= Cov(Xt − Xt−1 , Xt−k − Xt−k−1 )
= Cov(Xt , Xt−k ) + Cov(Xt , −Xt−k−1 ) + Cov(−Xt−1 , Xt−k ) + Cov(−Xt−1 , −Xt−k−1 )
= Cov(Xt , Xt−k ) − Cov(Xt , Xt−k−1 ) − Cov(Xt−1 , Xt−k ) + Cov(Xt−1 , Xt−k−1 )
= γk − γk+1 − γk−1 + γk = −γk+1 + 2γk − γk−1
only depends on time lag k. We conclude that {Wt } is stationary.

2.10. Let {Xt } be a zero-mean, unit-variance stationary process with

autocorrelation function ρk . Suppose that µt is a nonconstant function and
that σt is a positive-valued nonconstant function. The observed series is
formed as Yt = µt + σt Xt .
2.10.(a) Here we obtain the mean function
E(Yt ) = E(µt + σt Xt ) = µt + σt E(Xt ) = µt + 0 = µt
and the autocovariance function
Cov(Yt , Yt−k ) = Cov(µt + σt Xt , µt−k + σt−k Xt−k )
= Cov(σt Xt , σt−k Xt−k ) = σt σt−k Cov(Xt , Xt−k )
( √ )
= σt σt−k Corr(Xt , Xt−k ) V ar(Xt ) V ar(Xt−k )
( √ )
= σt σt−k ρk (1)(1) = σt σt−k ρk .

2.10.(b) It is easy to obtain V ar(Yt ) = V ar(µt + σt Xt ) = V ar(σt Xt ) =

σt2 V ar(Xt ) = σt2 (1) = σt2 , then the autocorrelation function
Cov(Yt , Yt−k )
Corr(Yt , Yt−k ) = √
V ar(Yt ) V ar(Yt−k )
σt σt−k ρk σt σt−k ρk
= √ = = ρk
2 2
σ σ σ t σ t−k
t t−k

only depends on time lag k. However, {Yt } is not stationary because the
mean function µt is not constant over time t.

2
 2.10.(c) If we change the setting such that µt = µ which is constant over
time t, then Wt = µ+σt Xt has constant mean function µ and autocovariance
function Cov(Wt , Wt−k ) = σt σt−k ρk by similar calculation above. Since the
autocovariance function depends on time t, then {Wt } is not stationary.
However, it has constant mean µ over time t.

2.13 Let Yt = et − θ(et−1 )2 , where et s are independent and identically

distributed (IID) random variables which follow N (0, σe2 ) distribution. Then
e2t /σe2 are IID random variables which follow chi-squared distribution with 1
degree of freedom with E(e2t /σe2 ) = 1 and V ar(e2t /σe2 ) = 2. In other words,
E(e2t ) = σe2 and V ar(e2t ) = 2σe4 .
2.13.(a) First we obtain the autocovariance function

Cov(Yt , Yt−k ) = Cov(et − θ(et−1 )2 , et−k − θ(et−k−1 )2 )

= Cov(et , et−k ) + Cov(et , −θ(et−k−1 )2 ) + Cov(−θ(et−1 )2 , et−k )
+Cov(−θ(et−1 )2 , −θ(et−k−1 )2 )
= Cov(et , et−k ) + 0 + Cov(−θ(et−1 )2 , et−k ) + Cov(−θ(et−1 )2 , −θ(et−k−1 )2 )
= Cov(et , et−k ) − θCov((et−1 )2 , et−k ) + θ2 Cov((et−1 )2 , (et−k−1 )2 ).

If k = 0,

Cov(Yt , Yt ) = Cov(et , et ) − θCov((et−1 )2 , et ) + θ2 Cov((et−1 )2 , (et−1 )2 )

= V ar(et ) − 0 + θ2 V ar(e2t−1 ) = σe2 + θ2 (2σe4 ) = σe2 + 2θ2 σe4 .

If k = 1,

Cov(Yt , Yt−1 ) = Cov(et , et−1 ) − θCov((et−1 )2 , et−1 ) + θ2 Cov((et−1 )2 , (et−2 )2 )

= 0 − θE{[(et−1 )2 − E((et−1 )2 )][et−1 − E(et−1 )]} + 0
= −θE{[(et−1 )2 − 1][et − 0]}
= −θE{(et )3 − et } = −θE[(et )3 ] + E(et ) = 0 − 0 = 0.

where E[(et )3 ] = 0 because N (0, σe2 ) distribution is symmetric.

If k > 1,

Cov(Yt , Yt−1 ) = Cov(et , et−k ) − θCov((et−1 )2 , et−k ) + θ2 Cov((et−1 )2 , (et−k−1 )2 )

= 0 + 0 + 0 = 0.

3
Then the autocovariance function
{ 2
σe + 2θ2 σe4 for k = 0
γk = Cov(Yt , Yt−k ) =
0 for k ≥ 1
which only depends on time lag k. And the autocorrelation function
{
1 for k = 0
ρk = .
0 for k ≥ 1

2.13.(b) Here we obtain the mean function

E(Yt ) = E(et − θ(et−1 )2 ) = E(et ) − θE((et )2 ) = 0 − θ = −θσe2

which is free of time. Since the autocovariance function depends only on

time lag k and the mean function is constant over time t, we conclude that
{Yt } is stationary.

2.14.(a) For Yt = θ0 + tet , we have the mean function

E(Yt ) = E(θ0 + tet ) = θ0 + tE(et ) = θ0 + 0 = θ0

which is free of time. And the autocovariance function

Cov(Yt , Yt−k ) = Cov(θ0 + tet , θ0 + (t − k)et−k ) = Cov(tet , (t − k)et−k )

= t(t − k)Cov(et , et−k )
{ 2 2
t σe for k = 0
=
0 for k ≥ 1

which depends on time t. Then we conclude that {Yt } is stationary.

2.14.(b) For Wt = ∇Yt = Yt − Yt−1 , first we expand Wt as

Wt = ∇Yt = Yt − Yt−1 = θ0 + tet − (θ0 + (t − 1)et−1 )

= tet − (t − 1)et−1

and we have the mean function

E(Wt ) = E(tet − (t − 1)et−1 ) = tE(et ) − (t − 1)E(et−1 ) = 0 − 0 = 0

4
which is free of time t. And the autocovariance function
Cov(Wt , Wt−k ) = Cov(tet − (t − 1)et−1 , (t − k)et−k − (t − k − 1)et−k−1 )
= Cov(tet , (t − k)et−k ) + Cov(tet , −(t − k − 1)et−k−1 )
+Cov(−(t − 1)et−1 , (t − k)et−k ) + Cov(−(t − 1)et−1 , −(t − k − 1)et−k−1 )
= t(t − k)Cov(et , et−k ) − t(t − k − 1)Cov(et , et−k−1 )
−(t − 1)(t − k)Cov(et−1 , et−k ) + (t − 1)(t − k − 1)Cov(et−1 , et−k−1 ).
If k = 0,
Cov(Wt , Wt )
= t(t)Cov(et , et ) − t(t − 1)Cov(et , et−1 )
−(t − 1)(t)Cov(et−1 , et ) + (t − 1)(t − 1)Cov(et−1 , et−1 )
= t2 V ar(et ) − 0 − 0 + (t − 1)2 V ar(et−1 )
= t2 σe2 + (t − 1)2 σe2 = (2t2 − 2t + 1)σe2 .
If k = 1,
Cov(Wt , Wt−1 )
= t(t − 1)Cov(et , et−1 ) − t(t − 2)Cov(et , et−2 )
−(t − 1)(t − 1)Cov(et−1 , et−1 ) + (t − 1)(t − 2)Cov(et−1 , et−2 )
= 0 − 0 − (t − 1)2 V ar(et−1 ) + 0 = −(t − 1)2 σe2 .
If k > 1,
Cov(Wt , Wt−k )
= t(t − k)Cov(et , et−k ) − t(t − k − 1)Cov(et , et−k−1 )
−(t − 1)(t − k)Cov(et−1 , et−k ) + (t − 1)(t − k − 1)Cov(et−1 , et−k−1 )
= 0 − 0 − 0 + 0 = 0.
Then we obtain the autocovariance function

 (2t2 − 2t + 1)σe2 for k = 0
Cov(Wt , Wt−k ) = −(t − 1)2 σe2 for k = 1

0 for k > 1
which depends on time t. So we conclude that {Wt } is not stationary.
2.14.(c) For Yt = et et−1 , we have the mean function
E(Yt ) = E(et et−1 ) = E(et )E(et−1 ) = 0(0) = 0

5
which is free of time t.
And the autocovariance function
Cov(Yt , Yt−k ) = Cov(et et−1 , et−k et−k−1 )
= E(et et−1 et−k et−k−1 ) − E(et et−1 )E(et−k et−k−1 ).
If k = 0,
Cov(Yt , Yt ) = E(et et−1 et et−1 ) − E(et et−1 )E(et et−1 )
= E(e2t e2t−1 ) − [E(et et−1 )]2
= E(e2t )E(e2t−1 ) − [E(et et−1 )]2
= σe2 σe2 − [E(et )E(et−1 )]2 = σe4 − 02
= σe4 .
since σe2 = V ar(et ) = E(e2t ) − [E(et )]2 = E(e2t ) − 0 = E(e2t ) for all time t.
If k = 1,
Cov(Yt , Yt−1 ) = E(et et−1 et−1 et−2 ) − E(et et−1 )E(et−1 et−2 )
= E(et e2t−1 et−2 ) − E(et et−1 )E(et−1 et−2 )
= E(et )E(e2t−1 )E(et−2 ) − E(et )E(et−1 )E(et−1 )E(et−2 )
= 0(σe2 )(0) − 0(0)(0)(0) = 0 − 0
= 0.
If k > 0,
Cov(Yt , Yt−k ) = E(et et−1 et−k et−k−1 ) − E(et et−1 )E(et−k et−k−1 )
= E(et )E(et−1 )E(et−k )E(et−k−1 ) − E(et )E(tet−1 )E(et−k )E(et−k−1 )
= 0.
Then we obtain the autocovariance function
{ 4
σe for k = 0
Cov(Yt , Yt−k ) =
0 for k ≥ 1
which only depends on time lag k.
From previous results, we know that {Yt } has constant mean function
and autocovariance function which only depends on time lag k. Then we
conclude that {Yt } is stationary.

6
 2.19. Let Y1 = θ0 + e1 , and then for t > 1, define Yt recursively by
Yt = θ0 + Yt−1 + et . Here θ0 is constant.
2.19.(a) Here we expand Yt for three times,

Yt = θ0 + Yt−1 + et = θ0 + (θ0 + Yt−2 + et−1 ) + et

= 2θ0 + Yt−2 + et−1 + et = 2θ0 + (θ0 + Yt−3 + et−2 ) + et−1 + et
= 3θ0 + Yt−3 + et−2 + et−1 + et .

Then we observe that after expanding t times, Yt can be rewrite as

Yt = tθ0 + et + · · · + e1 .

2.19.(b) Here we obtain the mean function

E(Yt ) = E(tθ0 + et + · · · e1 ) = tθ0 + E(et ) + · · · + E(e1 ) = tθ0 + 0 + · · · + 0 = tθ0 .

2.19.(c) The autocovariance function

Cov(Yt , Yt−k ) = Cov(tθ0 + et + · · · e1 , (t − k)θ0 + et−k + · · · e1 )

= Cov(et + · · · e1 , et−k + · · · e1 ).

For k = 0,

Cov(Yt , Yt ) = Cov(et + · · · e1 , et + · · · e1 ) = V ar(et + · · · e1 )

= V ar(et ) + · · · V ar(e1 ) = tσt .

For k ≥ 1,

Cov(Yt , Yt−k ) = Cov(et + · · · e1 , et−k + · · · e1 )

= Cov((et + · · · + et−k+1 ) + et−k + · · · + e1 , et−k + · · · + e1 )
= Cov(et + · · · + et−k+1 , et−k + · · · + e1 ) + Cov(et−k + · · · + e1 , et−k + · · · + e1 )
= 0 + V ar(et−k + · · · + e1 ) = V ar(et−k ) + · · · + V ar(e1 )
= (t − k)σe2 .

So we obtain the autocovariance function

Cov(Yt , Yt−k ) = (t − k)σe2 for k ≥ 0

which depends on time t and time lag k. Then we conclude that {Yt } is not
stationary.

7
 2.(a) First, we obtain the mean function
E(Yt ) = E(Z1 cos(wt) + Z2 sin(wt) + et )
= E(Z1 ) cos(wt) + E(Z2 ) sin(wt) + E(et )
= 0+0+0=0
which is constant over time t.
Then we divide the autocovariance function into three parts
Cov(Yt , Yt−k )
= Cov(Z1 cos(wt) + Z2 sin(wt) + et , Z1 cos(w(t − k)) + Z2 sin(w(t − k)) + et−k )
= Cov(Z1 cos(wt), Z1 cos(w(t − k)) + Z2 sin(w(t − k)) + et−k ) (1)
+Cov(Z2 sin(wt), Z1 cos(w(t − k)) + Z2 sin(w(t − k)) + et−k ) (2)
+Cov(et , Z1 cos(w(t − k)) + Z2 sin(w(t − k)) + et−k ). (3)
For part (1),
Cov(Z1 cos(wt), Z1 cos(w(t − k)) + Z2 sin(w(t − k)) + et−k )
= Cov(Z1 cos(wt), Z1 cos(w(t − k))) + Cov(Z1 cos(wt), Z2 sin(w(t − k)))
+Cov(Z1 cos(wt), et−k )
= cos(wt) cos(w(t − k))Cov(Z1 , Z1 ) + cos(wt) sin(w(t − k))Cov(Z1 , Z2 )
+ cos(wt)Cov(Z1 , et−k )
= cos(wt) cos(w(t − k))V ar(Z1 ) + 0 + 0 = cos(wt) cos(w(t − k)).
For part (2),
Cov(Z2 sin(wt), Z1 cos(w(t − k)) + Z2 sin(w(t − k)) + et−k )
= Cov(Z2 sin(wt), Z1 cos(w(t − k))) + Cov(Z2 sin(wt), Z2 sin(w(t − k)))
+Cov(Z2 sin(wt), +et−k )
= sin(wt) cos(w(t − k))Cov(Z2 , Z1 ) + sin(wt) sin(w(t − k))Cov(Z2 , Z2 )
+ sin(wt)Cov(Z2 , +et−k )
= 0 + sin(wt) sin(w(t − k))V ar(Z2 ) + 0 = sin(wt) sin(w(t − k)).
For part (3),
Cov(et , Z1 cos(w(t − k)) + Z2 sin(w(t − k)) + et−k )
= Cov(et , Z1 cos(w(t − k)) + Cov(et , Z2 sin(w(t − k))) + Cov(et , et−k )
= 0 + 0 + Cov(et , et−k ) = Cov(et , et−k ).

8
 Then we combine the results of all three parts and obtain the autocovari-
ance function
Cov(Yt , Yt−k ) = cos(wt) cos(w(t − k)) + sin(wt) sin(w(t − k)) + Cov(et , et−k )
= cos(wt − w(t − k)) = cos(wk) + Cov(et , et−k )
{
1 + σe2 for k = 0
=
cos(wk) for k ≥ 1.
According to calculations above, {Yt } has zero mean and autocovariance
function which is only depends on k. Then we conclude that {Yt } is station-
ary.
2.(b)
Please see Figure 1. There is trend in this graph. The observations
oscillate around zero between −4 to 4.
3
2
1
Trigonometric process

0
−1
−2
−3

0 50 100 150 200 250

Time

Figure 1: Time series plot for carbon dioxide data

The followings are R commands to make the graph above:

9
omega = 0.5
Z = rnorm(2,0.1)
e.t = rnorm(250,0,1)
Y.t = e.t*0
for (i in 1:length(e.t)){
Y.t[i] = Z[1]*cos(omega*i) + Z[2]*sin(omega*i) + e.t[i]}
plot(Y.t,ylab="Trigonometric process", xlab="Time", type="o")
2.(c) There is an increasing trend of the time series plot because of
adding the linear trend term β0 + β1 t to model. So we can see that the mean
function is not constant over time t and {Yt } appears not to be stationary.
115
110
Trig process with linear trend

105
100

0 50 100 150 200 250

Time

Figure 2: Time series plot for carbon dioxide data

The followings are R commands to make the graph above:

Y.tilde = e.t*0
for (i in 1:length(e.t)){
Y.tilde[i] = 100 + 0.05*i + Z[1]*cos(omega*i) + Z[2]*sin
(omega*i) + e.t[i]}

10
plot(Y.tilde, ylab="Trig process with linear trend", xlab="Time", type="o")

2.(d) The first differences process {∇Ỹt } looks to have trend. From
Problem 2.9, we know that this process must be stationary.
2
First differences of Y.tilde

0
−2

0 50 100 150 200 250

Time

Figure 3: Time series plot for carbon dioxide data

The followings are R commands to make the graph above:

diff.Y.tilde = diff(Y.tilde)
plot(diff.Y.tilde, ylab="First differences of Y.tilde", xlab="Time", type="o")

11
 3.(i) Here we give an example of a process {Yt } with constant mean
but variance that increases with time. So we consider Yt = tet with mean
function

E(Yt ) = E(tet ) = tE(et ) = t(0) = 0

which is constant. However, the variance

V ar(Yt ) = V ar(tet ) = t2 V ar(et ) = t2 σe2

increases with time.

3.(ii) Here we give an example of a stationary process whose autoco-

variance does not go to zero as time lag goes to infinity. So we consider
2
Yt = X + et where X has mean zero and variance σX . X and white noise
process {et } are independent. Then we have the mean function of {Yt }

E(Yt ) = E(X + et ) = E(X) + E(et ) = 0 + 0 = 0

which is constant over time t. The autocovariance function

Cov(Yt , Yt−k ) = Cov(X + et , X + et−k )

= Cov(X, X) + Cov(X, et−k ) + Cov(et , X) + Cov(et , et−k )
= V ar(X) + 0 + 0 + Cov(et , et−k ) = V ar(X) + Cov(et , et−k )
{ 2
σX + σe2 for k = 0
= 2
σX for k ≥ 1

which only depends on time lag k. So we conclude that {Yt } is stationary.

But the autocovariance function does not go to zero as time lag k goes to
infinity.

3.(iii) Here we give an example of a nonstationary process {Yt } whose

autocovariance depends only on time lag k. So we consider Yt = t + et with
mean function

E(Yt ) = E(t + et ) = t + E(et ) = t + 0 = t

12
which is not free of time t. Then {Yt } is not stationary. Furthermore, we
obtain the autocovariance function

Cov(Yt , Yt−k ) = Cov(t + et , (t − k) + et−k ) = Cov(et , et−k )

{ 2
σe for k = 0
=
0 for k > 0
which depends only on time lag k.

3.(iv) Here we give an example of a process {Yt } that has nonzero

autocorrelation only at lag k = 1. Consider Yt = et −et−1 with mean function

E(Yt ) = E(et − et−1 ) = E(et ) − E(et−1 ) = 0 − 0 = 0

which is free of time t. First, we obtain the autocovariance function


 2σe2 for k = 0
Cov(Yt , Yt−k ) = −σe2 for k = 1 ,

0 for k ̸= 1, k > 0

which is free of time t. Then we conclude that {Yt } is stationary. (The

calculation of autocorrelation function is similar to Problem 2.12.. Please
refer to Homework 1.) Then the autocorrelation function

 1 for k = 0
ρk = − 1 for k = 1
 2
0 for k ̸= 1, k > 0
has nonzero autocorrelation only at lag k = 1.

3.(v) Here we give an example of a nonstationary process {Yt } whose first

differences are stationary. Consider the random process {Yt } which is given
by 2.19.(c) with θ0 = 0. That is Yt = et + · · · + e1 . By similar calculation,
we know that {Yt } is not stationary. However, the first differences of {Yt },
defined by Wt = Yt − Yt−1 = et , is stationary since {et } is a white noise
process.

13