You are on page 1of 13

Continues

NATIONAL UNIVERSITY OF SINGAPORE



ST3233 APPLIED TIME SERIES ANALYSIS
(Semester 1 : AY 2013-2014)
Name of examiner: Dr Leontine Alkema

December 2013 - Time Allowed: 2 Hours

_________________________________________________________________

INSTRUCTIONS TO CANDIDATES

1. This examination paper contains SEVENTEEN (17) questions and
comprises THIRTEEN (13) printed pages.
2. Candidates must answer ALL questions. Each question carries 2 marks. A
total of 34 marks is available.
3. This is a CLOSED-BOOK exam. You may use the HELP SHEET (2 pages,
printed double-sided) that is provided to you.
4. Non-programmable calculators may be used.

5. Information specific to this exam:
a. Provide your answers in this question paper itself. Do not write in the
grey shaded areas.
b. Justify your answers to ensure full credit. Algebraically if possible,
succinctly (clear and to the point) otherwise. I will take points off for
incorrect or irrelevant statements.
c. Unless otherwise stated, information from the help sheet may be used
without proof.
d. Unless otherwise stated, et refers to normally distributed white noise;
et ~ N(0, e
2
) (independent).

6. Please write your matriculation number only in the space below. Do not
write your name.


Matriculation No: ___________________________

Seat No: ___________________


Questions Marks
scored
Questions Marks
scored
Total
1&2 8&9
3&4 10&11
5 12&13
6 14&15
7 16 (34)
17

2 ST3233
1. Suppose Y
t
follows an MA(1) model with = 0.9:
Y
t
= e
t
0.9e
t1
.
Is the model invertible? Show working and explain what it means if Y
t
is invertible.
2. Given the following MA(1) model:
Y
t
= e
t
+ 0.4e
t1
,
with autocorrelation function
1
= 0.345 (rounded to 3 digits),
0
= 1 and
k
= 0 for
k = 0, 1.
What is the partial autocorrelation function for lags 1, 2 and 3? Show working.
4
Continues
3 ST3233
3. Given the following AR(2) model
Y
t
= e
t
+ 0.6Y
t1
+ 0.2Y
t2
,
what is the autocorrelation function
k
for k = 1, 2, 3?
4. Write pseudo R-code to simulate a time series of length 100 from the MA(2) process
Y
t
= e
t
+ 0.8e
t1
0.3e
t2
,
where e
t
N(0, 1) (independent) without using any R built-in time series functions.
4
Continues
4 ST3233
5. The sample ACF and sample PACF for a time series on sales Y
t
of length 500 are provided
below. Do these functions provide information on what ARMA(p, q) model may be appro-
priate for modeling Y
t
? If yes, state p and q and explain your answer. If not, explain why
not.
5 10 15 20

0
.
4
0
.
0
0
.
2
0
.
4
0
.
6
Lag
A
C
F
5 10 15 20

0
.
4
0
.
0
0
.
2
0
.
4
Lag
P
a
r
t
i
a
l

A
C
F
2
Continues
5 ST3233
6. For a time series of interest Y
t
, the output below (from the function armasubsets) shows
the BICs for various models tted to Y
t
. Given the output, which model(s) can be considered
as candidate models for describing the time series? Explain your answer.
B
I
C
(
I
n
t
e
r
c
e
p
t
)
Y

l
a
g
1
Y

l
a
g
2
Y

l
a
g
3
Y

l
a
g
4
Y

l
a
g
5
Y

l
a
g
6
e
r
r
o
r

l
a
g
1
e
r
r
o
r

l
a
g
2
e
r
r
o
r

l
a
g
3
e
r
r
o
r

l
a
g
4
e
r
r
o
r

l
a
g
5
e
r
r
o
r

l
a
g
6
210
260
260
270
270
280
280
280
2
Continues
6 ST3233
7. An AR(1) model is tted to a time series Y
t
and test results based on the (standardized)
residuals are as follows:
> mod <- Arima(Y.t, order = c(1,0,0))
> summary(mod)
ARIMA(1,0,0) with non-zero mean
Coefficients:
ar1 intercept
0.5582 0.0113
s.e. 0.0262 0.0083
sigma^2 estimated as 0.01355: log likelihood=731.47
AIC=-1458.94 AICc=-1458.92 BIC=-1444.22
> sres.t <- rstandard(mod) # standardized residuals
> shapiro.test(sres.t)
Shapiro-Wilk normality test
data: sres.t
W = 0.9989, p-value = 0.7892
> LB.test(mod, lag = 16)
Box-Ljung test
data: residuals from mod
X-squared = 363.7266, df = 15, p-value < 2.2e-16
Given the information from the two tests above, does the AR(1) model seem appropriate for
the data? Explain your answer.
2
Continues
7 ST3233
8. Below is the output after tting an ARMA(1,1) model to a time series Y
t
of length 500,
Y
t
= Y
t1
+ e
t
e
t1
. State the 95% CI (condence interval) for AND interpret the
interval.
> mod <- arima(Y.t, order = c(1,0,1))
> summary(mod)
ARIMA(1,0,1) with non-zero mean
Coefficients:
ar1 ma1 intercept
0.6160 -0.3186 0.0123
s.e. 0.1036 0.1264 0.0080
9. Suppose you have tted an MA(1) model Y
t
= e
t
e
t1
to a short time series, e.g., of
length n = 25. Explain why it may be necessary to construct a bootstrapped 95% CI
(condence interval) for (instead of using the information from standard R-output directly
to construct a 95% CI) and describe the procedure (in words, not R-code) for constructing
such a bootstrapped CI.
4
Continues
8 ST3233
10. Give the specication of the tted time series model for Y
t
, including the specication of the
white noise terms, given the following R-output:
> mod <- arima(Y.t, order = c(1,0,1), seasonal = list(order = c(1,1,0), period = 12))
> summary(mod)
ARIMA(1,0,1)(1,1,0)[12]
Coefficients:
ar1 ma1 sar1
-0.0055 0.6941 -0.5028
s.e. 0.0440 0.0310 0.0277
sigma^2 estimated as 0.01801: log likelihood=580.36
AIC=-1154.72 AICc=-1154.67 BIC=-1135.13
11. For a stationary AR(1) process Y
t
= Y
t1
+e
t
with e
t
N(0,
2
e
) (independent), E(Y
t
) = 0,
V ar(Y
t
) =

2
e
1
2
and
k
=
k
, give the likelihood function if we would only observe the sum
Y
1
+ Y
4
. Show working.
4
Continues
9 ST3233
12. Suppose Y
t
is given by
(Y
t
1) = 0.1(Y
t1
1) + e
t
0.4e
t1
+ 0.1e
t2
with
2
e
= 1 and the most recent Y
t
s and e
t
s as displayed below:
t 95 96 97 98 99 100
Y
t
-0.3 2.4 1.5 2.8 0.7 0.6
e
t
-1.1 0.9 1.0 2.1 0.4 -0.5
Give the minimum mean square error forecast for Y
101
.
13. You are forecasting an annual time series Y
t
using an AR(1) model with E(Y
t
) = 10,
V ar(Y
t
) =

2
e
1
2
, = 0.5 and
e
= 1. Suppose Y
2012
= 15 (the last observation for the
year 2012). Will the 95% prediction interval (PI) for Y
2012+g
ever include 0 for some future
year 2012 + g, corresponding to lead time g? If yes, calculate g. If not, explain briey why
not.
4
Continues
10 ST3233
14. Suppose X
t
is an MA(1) process and Y
t
= 0.3X
t2
+ Z
t
, where Z
t
is an MA(2) process,
independent of X
t
. State the k for which
k
(X, Y ) = 0. Show (brief) working.
15. Suppose Z
t
= Z
t1
+ e
t
. Are the following two model specications for Y
t
equivalent?
Model 1: Y
t
= X
t
+ Z
t
.
Model 2: Y
t
X
t
= (Y
t1
X
t1
) + e
t
.
Show working.
4
Continues
11 ST3233
16. Suppose that time series X
t
can be modeled with a stationary AR(1) model
X
t
=

X
t1
+ e
t
,
and suppose that time series Y
t
is given by Y
t
=

10
k=10

k
X
t+k
+ Z
t
, where Z
t
is some
ARMA process, independent of X
t
, and the
k
s are unknown.
The plots below show the sample CCFs r
k
(X, Y ) and r
k
(

X,

Y ) for observed series of X
t
and
Y
t
of length 208, and for series

X
t
= X
t


X
t1
and

Y
t
= Y
t


Y
t1
.
Do these graphs suggest that one or more of the
k
s are non-zero? If yes, which
k
(s)? If
not, why not? Explain your answer.
4 2 0 2 4

0
.
6

0
.
4

0
.
2
0
.
0
Lag
C
C
F
(
X
,

Y
)
4 2 0 2 4

0
.
6

0
.
5

0
.
4

0
.
3

0
.
2

0
.
1
0
.
0
0
.
1
Lag
C
C
F
(
X
t
i
l
d
e
,

Y
t
i
l
d
e
)
2
Continues
12 ST3233
17. Suppose return r
t
follows an ARCH(1) model:
r
t
=
t|t1

t
,

2
t|t1
= + r
2
t1
,
where
t
is a time series of independently and identically distributed random variables with
E(
t
) = 0 and V ar(
t
) = 1, and
t
is independent of past returns. Suppose that 0 < < 1
such that E(r
2
t
) is constant (does not depend on t). What is E(r
2
t
)? Show working.
2
Continues
13 ST3233
Scrap Page
(please do not remove this page from the test packet)
End of Paper

You might also like