Time Series Analysis

Open Book Examination WS 2021/22, February 10th, 2022

Prof. Dr. Roxana Halbleib

Chair of Statistics and Econometrics
Institute of Economics
University of Freiburg

This exam contains 3 questions, each of which is worth 15 points. You have to answer all 3 questions
within 90 minutes.

You must answer in English.

Your answers need to be hand-written. Do your best to write legibly.

At the end of the examination time you have 15 minutes to scan/photograph each page that you
would like to submit as part of your overall response. Compile all your content into one document
so you can submit all your content as a single exam response document (as a PDF). Please ensure
that all pages are included as part of the PDF before you submit the document. Moreover, make
sure that the quality of the exam response scans/photographs is appropriate: they are sharp and
do not miss parts of the hand-written pages.

Exams or parts of exam responses which are missing or cannot be read with reasonable effort will
not be graded.

The successful uploading of your exam response PDF will be automatically confirmed in ILIAS.
In the ONLY case that you do not get this confirmation, please send the exam response PDF by
email to [email protected].

You are only allowed to work individually and without any unauthorised help. This specifically
forbids collaboration with and help by other persons as well as communication about the content
of the exam, during the examination period, with persons except for the examination staff. A
violation of these rules will result in failing the exam.

Please do not forget to state your student identification number on each page of your exam response
document!! Not doing so may have serious consequences for the correction eligibility and also for
the archiving of your exam responses.

Good Luck!
1
Problem 1: ARMA Models
Let

Yt = 0.3 + 0.4Yt−1 − 0.9εt−1 + 0.2εt−2 + εt , (1)

where εt is a white noise process.

a) Is the process for Yt covariance stationary? Explain your answer analytically. 1P

b) Is the process for Yt invertible? Explain your answer analytically. 2P

c) What is the order of the ARIMA(p,d,q) process in (1) in terms of p, d and q? 1P

d) Derive E[Yt ] and V[Yt ], given that V[εt ] = 0.3. 6P

e) Assume that you have a time series of T = 200 observations for Yt . You have that Y200 =
0.7, Y199 = 0.3 and ε200 = 0.08 and ε199 = −0.04.

i) Compute the 1-step and the 2-step ahead forecast of Yt . 3P

ii) Compute the forecast of Yt over the next two periods. 2P

2
Problem 2: Multivariate Time Series
Assume the bivariate Yt process given below:
   
0.2 0.8 b
Yt = + Y + εt , (2)
0.3 0.64 −0.8 t−1

 
0.3 σ1,2
where Yt = (Y1t , Y2t )0 and εt is a multivariate white noise process with V[εt ] =
σ1,2 0.7

a) Derive the range of values of b for the bivariate

√ process Yt to be covariance-stationary? Hint
1 i
1: When x < 0, x = − −x , where i = −1 is the imaginary number.
√ √ 5P

b) For which values of b and σ1,2 :

i) Does Y2 not Granger cause Y1 ? 1P

ii) Does Y1 not instantaneously cause Y2 ? 1P

Assume you estimate the bivariate process in Equation (2) based on T = 1000 observations. You
get the output in Figure 1.
Summary of Regression Results
------------------------------------------------------------
T 999 AIC 4.0188
BIC 4.0482 HQ 4.0300
------------------------------------------------------------
coeff. std.err. t-stat p-value
Results for equation 1
------------------------------------------------------------
const 0.2617 0.0293 8.9427 0.0000
L1.y1 0.7764 0.0208 37.4164 0.0000
L1.y2 0.0009 0.0145 0.0594 0.9526

Results for equation 2

------------------------------------------------------------
const 0.2704 0.0426 6.3518 0.0000
L1.y1 0.6578 0.0302 21.7920 0.0000
L1.y2 -0.7528 0.0212 -35.5901 0.0000

Covariance matrix of residuals

y1 y2
y1 0.3050 0.0876
y2 0.0876 0.6454

Figure 1: Estimation results.

c) Based on the output in Figure 1, what can you conclude about the Granger causality between
Y1 and Y2 ? Justify your answer. 1P

d) Based on the output in Figure 1, compute the 1 step ahead forecast at time T , i.e. YT +1|T
where YT = (0.32, 0.27)0 . 3P

e) You now apply a Cholesky-orthogonalized impulse response (IR) analysis and obtain the IR
graphs and their 95% confidence interval (CI) given in Figure 2.
3
 Figure 2: Cholesky-orthogonalized IR for 20 steps.

i) Interpret the Cholesky-orthogonalized IR functions plotted in Figure 2. 2P

ii) What problem of the non-othogonal IR does the Cholesky-orthogonalized IR solve? 1P
iii) What is the disadvantage of the Cholesky-orthogonalized IR analysis? 1P

4
Problem 3: Model Selection
Consider a time series Yt with t = 1, ..., 200, for which you obtain the line graph displayed in Figure
3.

Figure 3: Line graph of Yt .

a) By analysing the plot from Figure 3, what can you say about the covariance-stationarity of
Yt ? Explain your answer. 1P

You apply the Augmented Dickey Fuller (ADF) test accounting for an intercept on the d differences
of the time series Yt , i.e (1 − L)d Yt , where d = 0, 1, 2, 3 and obtain the results displayed in Table 1.

d Test-Statistic p-value
0 0.3300 0.9287
1 -46.7129 0.0000
2 -7.0422 0.0000

Table 1: Results of ADF test with intercept and with lags choosen by the Akaike Information
Criterion.

b) What are the null and the alternative hypotheses of the ADF test? 1P

c) Given the values in Table 1, what can you conclude about the order of integration of Yt at
5% significance level? Justify your answer. 1P
Consider now the series
Ỹt = (1 − L)d Yt
with d chosen such that Ỹt becomes covariance stationary. Figure 4 displays the sample autocor-
relation function (ACF) and the sample partial autocorrelation function (PACF) of Ỹt as well as
their 95% confidence interval (CI).
d) What can you say about the serial correlation of Ỹt ? Explain your answer. 3P
You apply all possible combinations of ARMA(p,q) models including an intercept for P = 0, 1, 2
and q = 0, 1, 2 (excluing the intercept only model) to fit Ỹt . In Table 2 you find the p-values of the
Ljung-Box test for lags 1 to 5 as well as the information criteria values for each model and Figure
5 shows the ACF and PACF graphs of the residuals as well as their 95% CI for each model.

5
 e) Based on Table 2 as well Figure 5, find the model that fits Ỹt best. Justify your choice by
including in your answer all values (test, criteria, p-values) and by discussing correspondingly
the plots in Figure 5 and the entries of Table 2. 6P

Figure 4: The series Ỹt .

ARMA(0,1) ARMA(0,2) ARMA(1,0) ARMA(1,1)

lag1 0.005 0.905 0.980 0.047
lag2 0.000 0.000 0.000 0.000
lag3 0.000 0.000 0.000 0.000
lag4 0.000 0.000 0.000 0.000
lag5 0.000 0.000 0.000 0.000
AIC 4.306 3.800 4.648 4.276
BIC 4.355 3.866 4.698 4.342
HQ 4.326 3.827 4.668 4.303
(a) Ljung Box Test p-values for lags 1 to 5 and IC values for ARMA(0,1), ARMA(0,2), ARMA(1,0) and
ARMA(1,1).
ARMA(1,2) ARMA(2,0) ARMA(2,1) ARMA(2,2)
lag1 0.937 0.574 0.970 0.989
lag2 0.000 0.781 0.905 0.985
lag3 0.000 0.884 0.931 0.964
lag4 0.000 0.957 0.979 0.990
lag5 0.000 0.981 0.989 0.995
AIC 3.810 2.852 2.861 2.868
BIC 3.892 2.919 2.944 2.968
HQ 3.843 2.879 2.894 2.909
(b) Ljung Box Test p-values for lags 1 to 5 and IC values for ARMA(1,2), ARMA(2,0), ARMA(2,1) and
ARMA(2,2).

Table 2: Ljung Box Test statistic and p-values for lags 1 to 5 and IC values

6
(a) The ARM A(0, 1) residuals. (b) The ARM A(0, 2) residuals.

(c) The ARM A(1, 0) residuals. (d) The ARM A(1, 1) residuals.

(e) The ARM A(1, 2) residuals. (f) The ARM A(2, 0) residuals.

(g) The ARM A(2, 1) residuals. 7 (h) The ARM A(2, 2) residuals.

Figure 5: Residual sample ACF and PACF for the estimated models.
 f) Based on your result for the order of integration in c) and your result for the best choice of
a model at point e), compute the 1 step ahead forecast of Yt at t = 200, i.e. Y201|200 . You
have that Y200 = 12.30, Y199 = 11.93, Y198 = 12.37 and ε̃ˆ200 = −1.17, ε̃ˆ199 = −0.30 and
ε̃ˆ198 = −0.89 where ε̃ˆt is the residual of the model fitted to Ỹt at time t. The estimated
parameters of each model for Ỹt are given in Table 3 below. 3P

c φ1 φ2 θ1 θ2
ARMA(0,1) -0.12 -0.85
(0.02) (0.04)
ARMA(0,2) -0.12 -0.07 -0.76
(0.02) (0.04) (0.03)
ARMA(1,0) -0.13 -0.01
(0.17) (0.07)
ARMA(1,1) -0.12 0.22 -0.89
(0.02) (0.08) (0.03)
ARMA(1,2) -0.12 0.01 -0.08 -0.76
(0.02) (0.08) (0.04) (0.03)
ARMA(2,0) -0.13 -0.01 -0.91
(0.04) (0.03) (0.03)
ARMA(2,1) -0.13 0.01 -0.91 -0.04
(0.04) (0.03) (0.03) (0.07)
ARMA(2,2) -0.13 0.01 -0.92 -0.05 0.05
(0.04) (0.03) (0.03) (0.08) (0.08)

Table 3: Estimation results for the different models. Standard errors in parenthesis.

∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗

Please do not forget to upload your declaration of independent work, filled in and signed!

∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗∗