1

Copyright reserved

University of Pretoria
Department of Statistics
WST 321 & STK 320
Time Series Analysis

Semester Test 2 9 October 2018

Examiners: Dr. J.T. Ferreira
Dr. D.A. Burger
Reading time: 5 minutes
Time (WST 321): 1 hour 5 minutes
Time (STK 320): 1 hour 30 minutes
Pages: 7
Questions: 6
Marks: 36 (WST 321), 50 (STK 320)

Read and follow the instructions on the next page.

Note that STK 320 has a separate, additional paper to be completed (14 marks).
 2

INSTRUCTIONS:

Answer Questions 1 to 6 (WST 321) in one book.

STK 320 students: answer Dr. Burger’s section on the printed test.
Please number your manuscript book with the question numbers you com-
pleted.
Give su¢ cient explanations for each step - no marks will be awarded without
the necessary explanations.
Make sure to clearly de…ne and indicate any and all notation that you use.
The usual examination and test rules and regulations of the University of
Pretoria are valid for this assessment.
Make sure to hand this test in after you wrote it.
Use three decimal places where necessary.
Assume that fat g is an independent white noise process with E (at ) = 0 and
2
var (at ) = a.
 3

Answer Questions 1 to 6 in one book

1. Identify a parsimonius ARIM A(p; d; q) model for the following time series process by deter-
mining the values of p, d, and q. Make sure to show all your steps.

Zt = 2Zt 1 Zt 2 + at :

(4)

2. Consider the following time series model where t = 1; 2; 3; :::; n:

Zt = 0 + 1t + at

Suppose the coe¢ cients of this model are estimated using ordinary least squares (OLS).

(a) Write down E (Zt ) of this model, and derive ^ 0 , the OLS estimator for 0. (3)
(b) Assuming that ^ 1 is an unbiased estimator for 1, prove that ^ 0 is an unbiased estimator
for 0 . (2)

3. Suppose fWt g is a stationary M A(1) process, given by

Wt = 0 + at at 1 :

Let Wt = Zt , where denotes the …rst di¤erence operator. Suppose that the usual
conditions and assumptions for integrated time series are valid.

(a) Write down the process of Zt , and identify the process fully. (2)
tPm
(b) Derive the general linear process representation of Zt by using Zt = Wt j . (2)
j=0

(c) Derive the variance at time t of Zt . (2)

4. Suppose you are requested by your good friend (Johan) to decide between …tting an AR(1)
or an AR(2) model to given data, consisting of 24 observations. The sample autocorrelations
at di¤erent lags k for the data is obtained and given in the following table:

k 1 2 3 4 5
rk 0:712 0:684 0:201 0:107 0:004

(a) Write down the relevant null- and alternative hypothesis that would be used to decide
between the two suggested models. (2)
(b) Construct an approximate 95% con…dence interval to test the hypothesis. Make sure
to show all your steps and reach a …nal conclusion regarding the hypothesis, as well as
r r2
suggesting a …nal model (hint: ^ 11 = r1 and ^ 22 = 12 r21 ). (3)
1
 4

5. We’ve seen that many observed time series have a unit root. The Dickey-Fuller (DF) test was
developed to test whether a given time series potentially has a unit root or not. Use the SAS
output given for Question 5 to answer the questions where relevant.

(a) Write down the null- and alternative hypothesis for the DF test, and de…ne all notation
used. (2)
(b) The DF test relies on an AR(1) model as the underlying data generating process (d.g.p).
The Augmented Dickey-Fuller (ADF) test was then developed to account for other d.g.ps
with additional lags. Consider an AR(2) model as candidate for the d.g.p for a zero-mean
process:
Z t = 1 Z t 1 + 2 Z t 2 + at :
Derive an augmented speci…cation which is used in the ADF test to test for the existence
of a unit root. Remember to clearly specify all notation used and introduced. (3)
(c) Your good friend (Johan) made some mistakes in the SAS code below when he attempted
to use the ADF test for the underlying AR(2) d.g.p as in Question 5 (b). Write down
the three mistakes which would not give the output as given (note: the name of the
dataset used is correct, as well as the procedure being used...). (3)
proc arima data=datata plots=none;
estimate var=zt nlag=2 stationarity=(adf=(0,1,2,3,4,5,6));
run;
(d) Your good friend (Johan) has asked you to use the ADF test to determine whether there
is a unit root for data he provided. Use the output for Question 5 to determine whether
the given time series has a unit root or not, using the studentised test statistic. State
which case you are using by consulting the graph of the data. Assume an underlying
d.g.p in the form of an AR(2) model. Give the test statistic and the associated p-value,
and reach a conclusion using = 0:01. (3)
 5

6. Suppose the method of maximum likelihood was used to estimate parameters of an AR(2)
model from 200 observations, after your good friend (Johan) provided you with the data.
Answer the questions that follow.

(a) Use the SAS output given for Question 6 to obtain the following values. Redraw this
table in your manuscript book in order to complete it. (4)

Estimate Value
^ ?
1
^ ?
2

^ ?
^ 2a ?

(b) Construct and give an approximate 95% con…dence interval which may be used to test
inference regarding the parameter 2 . Do this without consulting the SAS code
provided. (2)

Total: 36
 6

Question 5 additional information

 7

Question 6 additional information

 8

Memorandum: WST 321 and STK 320 Semester Test 2 - 9 October 2018

1. Consider Zt = 2Zt 1 Zt 2 +at . Thus we can write (1 2L + L2 ) Zt = at . The characteristic

equation is given by (1 2x + x2 ) = (1 x)2 = 0, and thus we have two unit roots, both
equal to 1: x1 = x2 = 1.
The process is thus non-stationary and d = 2.
If we then consider Wt = (1 L)d Zt = (1 L)2 Zt , then Wt = at , and thus p = q = 0.
Thus we have an ARIM A(0; 2; 0) model.

2.

(a) E (Zt ) is given by

E (Zt ) = 0 + 1 t:

The sum of squared errors are given by

P
n
Q = [Zt E (Zt )]2
t=1
Pn
2
= [Zt 0 1 t] :
t=1

The derivative for 0 is given by

dQ P
n
= 2 [Zt 0 1 t] :
d 0 t=1

Setting this derivative equal to 0:

^0 = Z 1^ P n
t:
n 1 t=1

(b) See that

1^ Pn
E ^0 = E Z t
n 1 t=1
1P n 1^ P n
= E Zt t
n t=1 n 1 t=1
1 P
n P
n
= E (Zt ) E ^ 1 t
n t=1 t=1
1 P n P
n
= ( + 1 t) t
n t=1 0 1
t=1
1 Pn Pn
= n 0+ 1 t 1 t
n t=1 t=1
= 0:
 9

3.

(a) Since Wt = Zt , we have that

Zt = 0 + at at 1

and thus
Zt = 0 + Zt 1 + at at 1

which is an IM A(1; 1) process, or a non-stationary ARM A(1; 1) process.

(b) See that
tPm
Zt = Wt j
j=0
tPm
= ( 0 + at j at j 1)
j=0
= (t m + 1) 0 + (at at 1 ) + (at 1 at 2 ) + ::: + (am am 1 )
= (t m + 1) 0 + at + (1 )at 1 + (1 ) at 2 + ::: + (1 ) am am 1 :

(c) The variance of this process at time t is given by

t;t = var (Zt )

= var ((t m + 1) 0 + at + (1 )at 1 + (1 ) at 2 + ::: + (1 ) am am 1 )
2 2 2
= 1 + (t m) (1 ) + a:

4.

(a) H0 : kk = 0 versus H1 : kk 6= 0.
q q
(b) An approximate 95% con…dence interval is given by 2 n1 ; 2 n1 (from formula
q q
1 1
sheet). Thus 2 24 ; 2 24 = ( 0:408; 0:408).
r r2 0:7122
Now: ^ 11 = r1 and ^ 22 = 12 r21 , thus ^ 11 = 0:712 and ^ 22 = 0:684
1 0:7122
= 0:359.
1
^ 11 lies outside the con…dence interval and is thus considered to be signi…cantly di¤erent
from 0, but ^ 22 lies within the con…dence interval. Therefore it seems to be not signi…-
cantly di¤erent from 0, and we conclude that an AR(1) model seems more appropriate
than an AR(2) model for this data.
 10

5.

(a) H0 : = 1 versus H1 : < 1.

(b) See that
Zt = 1 Zt 1 + 2 Zt 2 + at
as given. Subtract Zt 1 from both sides:
Zt = 1 Zt 1 + 2 Zt 2 Zt 1 + at
= ( 1 1) Zt 1 + 2 Zt 2 + at
= ( 1 1) Zt 1 + 2 Zt 2 + 2 Zt 1 2 Zt 1 + at
= ( 1 + 2 1) Zt 1 2 (Zt 1 Z t 2 ) + at
= ( 1 + 2 1) Zt 1 2 ( Z t 1 ) + at
= ( 1) Zt 1 2 ( Z t 1 ) + at :
If = 1 the …rst di¤erence of Zt is a stationary AR(1) process, but Zt has a unit root.
This augmented speci…cation is used to test H0 : = 1 versus H1 : < 1.
(c) identify var=zt nlag= 6 stationarity=(adf=( 0,1,2 ));
(d) For an underlying AR(2) d.g.p., we have to pick the line with lag 1.
Case: Trend
= 2:47 with p-value = 0:3413.
OR
Case: Single Mean
= 1:32 with p-value = 0:6179.
In both cases, we do not reject the null hypothesis and conclude that the series does
have a unit root.

6.

Estimate Value
^1 0:795
(a) ^2 0:118
^ 121:796
^ 2a 3:487
2
(b) From formula sheet: var ^ 2 1 2
. Thus an approximate 95% con…dence interval
r r n
is given by 2 var ^ 2 ; 2 var ^ 2 which gives ( 0:1404; 0:1404), which may
be used to test hypotheses regarding 2.

END