FORECASTING THE GROSS DOMESTIC PRODUCT (GDP) OF MALAYSIA

BRENDA BOPULAS

This project is submitted in partial fulfillment of

the requirement for the degree of Bachelor of Economics with Honours
(International Economics)

FacultY of Economics and Business

UNlVERSITI MALAYSIA SARAWAK
2011

ABSTRAK

MERAMAL KELUARAN DALAM NEGERI KASAR (KNDK) DI

MALAYSIA

Oleh

BRENDA BOPULAS

Kajian ini bercadang untuk mengkaji penentu keluaran dalam Negeri kasar (KNDK)
di Malaysia dan kemudian menggunakannya untuk meramal KNDK Malaysia. Ujian
empirikal yang digunakan termasuk ujian kepegunan, ujian kopengamiran Johansen
dan pekali korelasi. Model yang digunakan untuk meramal KNDK ialah model
Autoregressive Integrated Moving Average (ARIMA), model asas dan model
Random Walk. Keputusan daripada kajian ini mengesahkan bahawa bekalan wang,
pengeluaran perindustrian, eksports dan perbelanjaan penggunaan isi rumah
mempunyai perkaitan yang kuat dengan KNDK. Seterusnya, kajian ini juga
mengesahkan bahawa untuk meramal KNDK tepat, semua pembolehubah yang
mempunyai perkaitan dengan KNDK perlulah dimasukkan kerana persamaan yang
hanya ada satu pembolehubah atau yang hanya melibat KNDK sahaja akan lebih
kepada menghasilkan ramalan yang kurang tepat.

ABSTRACT

....

FORECASTING THE GROSS DOMESTIC PRODUCT (GDP) OF

MALAYSIA

By

BRENDA BOPULAS

This study intends to examine the determinant of Gross Domestic Product (GDP) of
Malaysia and use them to forecast the GDP of Malaysia. The empirical test that is
used in this study includes unit root test, Johansen cointegration test and correlation
test. The models that are employed are Autoregressive Integrated Moving Average
(ARIMA), fundamental models and Random Walk ModeL The results state that
money supply, industrial production, exports and household consumptions have
strong relationship with

GDP and in order to forecast the GDP of Malaysia, all

variables that are important and could impact the GDP should be included. Single
equation model tends to produce less accurate forecast.

Acknowledgement

I would like to grab the opportunity to offer special thank to the organization and all
the people involve that had assisted me in complementing this paper.

First and foremost, I would like to say thank you to my university, University
Malaysia Sarawak (UNIMAS) for their support and also effort in ensuring that all of
the third year students would be able to take their final year project as it is one of the
prerequisites in order to enable the students to qualifY for their graduation. I would
also like to thank my faculty, Faculty of Economics and Business (FEB) for all their
support and also the resources that they had provided in order for me to successfully
complete my paper.

Secondly, a warm thank you also to my supervisor, Associate Professor Dr Venus

Khim Sen-Liew for the time and patient that he had invested in supervising me all
the way until the completion of this paper is made possible. This paper would not be
able to be completed without his guidance and advices.

Not forgetting also the lecturers of Faculty of Economics and Business (FEB) that
had teaches me all the fundamental knowledge and concept of economics from
scratches as it was my first, time learning economics. The knowledge that I learn all
the while had help me in completing my paper.

Lastly, I would also like to thank all my friends, course mates and family that always
by my side to motivate me and offer encouraging words to me that help me
overcome all the difficulties and tension during the process of doing this paper,
without their motivation and trust, I would not be able to complete this paper.

TABLE OF CONTENTS

LIST OF TABLES ................................................................................................ viii-ix

LIST OF FIGURES ...................................................................................................... x

CHAPTER 1: INTRODUCTION

1.0

Introduction...................................................................................................... 1-2

1.1

Concept of Study ............................................................................................. 2-3

1.1.1 Importance of Economic Forecasting ....................................................... .3

1.1.1.1 Individual. ..................................................................................... 3

1.1.1.2 Business ........................................................................................ 3

1.1.1.3 Financial Institution ..................................................................... .4

1.1.1.4 Government. ................................................................................. 4

1.1.2 Forecasting Gross Domestic Product (GOP) in Malaysia..................... .4-5

1.2

Background of Study ......................................................................................... 5

1.2.1 History and Governance of Malaysia..................................................... 5-7

1.2.2 Geography .............................................................................................. 7-8

1.2.3 Economy...............................................................................................7-12

1.3

Motivation of Study..................................................................................... 12-13

1.4

Problem Statement. ...................................................................................... 13-15

1.5

Objective of Study............................................................................................ 15

1.5.1 General Objective .................................................................................... 15

1.5.2 Specific Objective.................................................................................... 15

1.6

Significance of Study .................................................................................. 15-16

1.7

Structure of Study............................................................................................. 17

CHAPTER 2: LITERATURE REVIEWS

2.0

Introduction.................................................................................................. 18-19

2.1

Theoretical Framework..................................................................................... 19

2.1.1 Stock Market. ..................................................................................... 19-20

2.1.2 Real Activity ............................................................................................20

2.1.3 Money Supply..........................................................................................21

2.1.4 Exchange Rate ....................................................................................21-22

2.1.5 Interest Rate ........................................................................................22-23

ii

2.1.6 Trade........................................................................................................23

2.1.7 Consumption Expenditure ................................................................. .23-24

2.2

Empirical Testing Procedure .............................................................................24

2.2.1 Specification of Models......................................................................24-27

2.2.2 Forecasting Models ..................................................................................27

2.2.2.1 Vector Autoregressive (VAR) ModeL ................................ .28-29

2.2.2.2 Dynamic Stochastic General Equilibrium (DSGE) modeL ...... .30

2.2.2.3 Dynamic Factor Model.. ....................................................... 30-32

2.2.2.4 Univariate Autoregressive Integrated Moving Average

(ARIMA) ..............................................................................32-33

2.2.3 Empirical Method ....................................................................................33

2.2.3.1 Stationary test. .......................................................................33-34

2.2.3.2 Johansen Multivariate Cointegration Test .............................34-35

2.2.4 Forecast Criteria......................................................................................35

2.2.4.1 Information Criteria................................................................... .36

2.2.4.2 Root Mean Square Error (RMSE) ........................................36-37

2.2.4.3 Mean Absolute Percentage Error (MAPE) ........................... .37-38

iii

2.2.4.4 Encompassing tests .....................................................................38

2.3

Empirical Evidence......................................................................................39-43

2.4

Concluding Remarks ...................................................................................43-44

CHAPTER 3: METHODOLOGY

3.0

Introduction..................................................................................................57-58

3.1

Model and Data Description........................................................................ 58-61

3.2

Empirical Testing Procedure.............................................................................61

3.2.1 Unit Root Tests ........................................................................................62

3.2.1.1 Augmented Dickey-Fuller (ADF) .........................................62-63

3.2.1.2 Phillips-Perron (PP) ...............................................................63-64

,
I

3.2.2 Cointegration Test. .............................................................................64-65

3.2.3 Forecasting Model. ..................................................................................65

3.2.3.1 ARIMA Model. ....................................................................65-66

3.2.3.2 Fundamental Models .................................................................66

3.2.3.3 Random Walk ModeL ...............................................................66

3.2.4 Model Specification and Diagnostic Checking .......................................67

iv

1I
3.2.4.1 Normality Test. ...........................................................................67

3.2.4.2 Correlation Test.. ..................................................................67-70

3.2.6 Forecast Evaluation Criteria....................................................................70

3.2.6.1 Root Mean Square Error (RMSE) ..............................................70

3.2.6.2 Mean Absolute Percentage Error (MAPE) ............................70-71

CHAPTER 4: RESULTS AND DISCUSSION

4.0

Introduction..................... ,............................................................................ 71-72

4.1

Unit Root Test ...................................................................................................73

4.1.1 Unit Root Test Result. ........................................................................73-79

4.2

ARIMA Model ..................................................................................................80

4.2.1 ARIMA Model Forecasting Accuracy ................................................80-83

4.2.2 Comparison between yearly and quarterly ARIMA model.. ............. 83-84

4.3

Fundamental Models ...................................................................................84-85

4.3.1 Model 1.............................................................................................86-87

4.3.2 Model 2 .............................................................................................88-89

4.3.3 Model 3.............................................................................................89-93

,...

4.3.4 Performance of the Fundamentals Models ..............................................93

4.4 Random Walk Model. .......................................................................................94

4.4.1 Forecast Accuracy of Random Walk Model.. ......................................... 94

4.5 Comparisons between ARIMA, Fundamental models and Random Walk

Model................................................................................................................95

CHAPTER 5: CONCLUSION AND RECOMMENDATION

5.0 Introduction......................................................................................................96

5.1 Summary of Finding....................................................................................97-99

5.2 Policy Implication and Suggestion...................................................................99

5.2.1 Public Expenditure................................................................................ 100

5.2.2 Foreign Market Access .......................................................................... 101

5.2.3 Monetary Policy.................................................................................... 102

5.3 Limitation........................................................................................................ 103

5.4 Recommendation for Future Studies ............................................................... 103

5.5 Contribution to Study...................................................................................... 104

5.6 Concluding Remark ................................................................................. 104-105

vi

REFERENCES................................................................................................ 106- 112

APPENDIX

vii

LIST OF TABLE

Table 2.1: Summary of Literature Reviews ......................................................... .45-56

Table 4.1: ADF results for 1970 to 2010 .................................................................... 76

Table 4.2: PP results for 1970 to 2010 ....................................................................... 77

Table 4.3: ADF results for 1970 to 2008 ................................................................... 78

Table 4.4: PP results for results for 1970 to 2008 ...................................................... 79

Table 4.5: Forecasting Performance of ARlMA models with constant (Yearly) ....... 82

Table 4.6: Forecasting Performance of ARIMA models without constant (Yearly) ..83

l
Table 4.7: Johansen Multivariate Co integration test for Model 1.............................. 87

': .

,.
{

Table 4.8: Johansen Multivariate Cointegration test for Model 2 .............................. 88

Table 4.9: Jarque-Bera Normality Test ...................................................................... 90

Table 4.10: Pearson Correlation Test for 1970 to 2010 .............................................90

Table 4.11: Pearson Correlation Test for 1970 to 2008 ............................................. 90

Table 4.12: Johansen Multivariate Cointegration test for Model 3 ............................92

Table 4.13: Forecasting Performance of Fundamental models ..................................93

viii

Table 4.14: Forecasting Performance of Random Walk ModeL............................... 94

Table 4.15: Forecasting Performance of ARIMA, Fundamental Model and Random

Walk Model ............................................................................................. 95

Table AI: PP results for results for 1991QI-201OQ4................................. Appendix A

Table A2: PP results for results for 1991QI-2008Q4 ................................ Appendix A

Table Bl: Forecasting Performance of ARMA models with constant

(Quarterly) .................................................................................. Appendix B

ix

LIST OF FIGURE

Figure 1: Map of Malaysia............................................ .... ...... .. ...................... 8

Figure 2: Trend of GOP in Malaysia from 1980-2009................................................ 9

,',

.j

"

,....

Chapter One

Introduction
1.0

Introduction

This study attempts to forecast the Gross Domestic Product (GDP) of

Malaysia. Forecasting is a process to estimate the future value and it could be done
by a lot of agents which include individual, businesses, financial institution and even
government. The objectives in this study are to determine the determinant of GDP, to
forecast the GDP and later to evaluate the forecast accuracy. The factors considered
include money supply in which according to theory is positively related to GDP
(Arnold, 2008), interest rate which has a theoretically negative relation with GDP
(Obamuyi, 2009), exchange rate which is negatively related (Rodrik, 2008),
household consumption expenditure, industrial production and also exports which
rt

these three factors according to theory are all positively related the (Kogid, Mulok &

!I

Lim, 2010; Fama, 1981; Fama, 1982; Keynes, 1936).

To forecast Malaysia GDP with good accuracy Autoregressive Integrated

Moving Average (ARIMA) time series models, fundamental models and random
walk model are estimated. Then, the best forecasting model that could accurately
forecast the GDP will be identified.

To preview this study results, the major finding is that broad money, household
consumption expenditure, industrial production and exports are very important to the
economy growth of Malaysia. Moreover, fundamental models performed the best

'P

,
i

from ARIMA time series model and random walk model. Random walk model

..

performed the worst among the three models. However, this is against the spirit of
parsimony that state simple model specification is better (Gilbert, 1995).

Chapter 1 is organized as below. Section 1.1 discuss briefly on the concept of

study. Then, Section 1.2 gives a brief overview on the background of study. Section
1.3 discusses the significant of the study. This is followed by Section 1.4 which
states the motivation of doing the study. After that, Section 1.5 will discuss the
problem statement in the study and Section 1.6 lists out the objectives of doing this
study. Finally, Section 1.7 provides the organisation of the study.

1.1

Concept of Study

Forecasting in general is the process of estimating the future value of a

variable. Forecasting is an extremely complex activity that could influence the
setting up of an organization. Forecasts of macroeconomic variables are also crucial

"j

i"
~,

~'

to many agents of the economy and that include Central Bank, commercial banks,
investors, speculators, government, policy maker and individual which includes
household. As forecast it is important to most of the agents of the economy thus
there are a number of studies done to forecast the economic performance. This
includes studies to forecast the inflation (Serrato, 2006; Mehrotra & Sanchez-Fung,
2008; Beechey & Osterholm, 2010), GDP (Krajewski, 2009; Mittnik & Zadrozny,
2004; Lu, 2009; Gupta, 2006), national accounts (Angeline, Banbura & Runstler,
2008), interest rate (Bidarkota, 1998; Fletcher & Gulley, 1996; Byers & Nowman,
1998)), exchange rate (Grossmann & McMillan, 2010; Carriero, Kapetanios &

,"

Marcellino, 2009) and some other kinds of forecasts. The importances of economic
forecasting in general are discussed below:

1.1.1 Importance of Economic Forecasting

1.1.1.1

Individual

For individuals, economic forecast helps them connect to the economic affairs
as they will get brief and fair ideas on how the things are going to be in the future.
Forecasting could help people to be in charge of the economic affairs and make
better decision that may help them to avoid losses.

1.1.1.2

Business

As the business environment is constantly changing, forecast could help

","

,,'

company to foresee the future which is important in enabling the management to

change operation at the right time in order to reap greatest benefit. In addition,
forecasting could also avoid the company from losses by making a proper decisions
based on relevant information. This is done by forecasting the economy and market
to establish the pattern of the market, the size and its growth potential. Besides,
forecasting could give them a brief idea on the kinds of government policy that
would be used as businesses performance are also sometimes ties to the policy
introduced by the government.

,'{

,I

r
1.1.1.3

Financial Institution

Forecasting could help anchor the expectation of the firms and households
which could help financial institution to become more effective in fulfilling the
demand of the individual and businesses. Besides, forecast publications could also
prevent them from making wrong investment that could result in great losses.

1.1.1.4

Government

Forecasting also plays an important role especially for Government strategic

planning and when it is required to do certain long term projects and evaluation of
the economy. Most ofthe time, forecasting is carried out when there is a need to seek
for aid to decision making and certain planning in the future. Especially when
Government intend to introduce new economic policy, it is important to know the
trend for the country economy in order to make sure that the policy introduced is
suitable with the country's economy situation.

1.1.2 Forecasting Gross Domestic Product (GDP) in Malaysia

The aim of this study is to forecast the GDP of Malaysia. There are a lot of
agencies in Malaysia that forecast GDP. These agencies include Malaysia Institute of
Economic Research (MIER), Bank Negara Malaysia (BNM), Amanah Mutual
Berhad (AMB) and other commercial banks. Producing accurate forecast in
Malaysia is important.

This is important as the forecasted value of the Gross Domestic Product can
give an overview on how the economy would be behaving in the future and enable
4

.'i

1 '

i. ~

the policymaker to come up with policies that would suit the economic condition in
the future. This eventually could prevent Malaysia to encounter any kind of crisis as
we would already have an overview on the country economy and policy could be
implemented earlier as to counter these problems. For example the financial crisis in
1997 to 1998 where with forecasting the effect of debt if Malaysia borrow from
International Monetary Fund (IMF) it is found out that Malaysia would not be able
to finish paying it for a very long time as the interest is very high thus capital control
and fixing the currency is adopted. This action shows that by forecasting the effect of
a policy then unnecessary policy could be prevented from being implemented.

As for individuals, forecasting GDP could expose to the individual the trend of
the GDP which could represent how well and how the country income is doing. An
early sign of a decreasing trend could signal an individual to be more careful in their
investment or some other activities that they are doing. For example, if each
individual has been exposed to the GDP forecast publication then they would know
that Malaysia is not doing so well between 1997 and 1998 thus can tell them to sell
their stocks so that they will not encounter high losses during the financial crisis.

1.2

Background of Study

1.2.1 History and Governance of Malaysia

Before the Malay Peninsula gained its independence in 1957, Malaysia was
initially ruled by Great Britain in the late 18th and 19th centuries. Later, Japan began
to occupy Malaysia from 1942 to 1945. Malaysia'S first prime minister, Tunku
Abdul Rahman Putra (1957-1970) has brought Malaysia from colonialism to
5

independence and he also proposed the idea of Malaysia. Malaysia was officially
formed in 1963 when the British colonies in Singapore and the East Malaysian states
of Sabah and Sarawak joined the Federation. In the beginning of the several years,
Malaysia was marred by a communist insurgency, confrontation with Indonesia, The
Philippines' claim to Sabah and Singapore withdrawal in 1965. Malaysia's second
prime minister, Tun Abdul Razak Bin Dato'Hussein (1970-1976) launched New
Economic Policy (NEP) in 1971 which consists of two basic goals which are to
eradicate poverty and eradicate identification of economic function with race.

Malaysia third prime minister, Tun Hussein Onn (1976-1981) stressed on the
issue of unity through policies aimed at rectifying economic imbalances between
communities which result in the launching of National Unit Trust Scheme in 1981.
He also took a serious consideration in the concept of Rukun Tetangga and against
drug. Malaysia's fourth prime minister, Tun Dr. Mahathir bin Mohamad (1981
2003) also had successful diversified Malaysia economy from dependence on
exports of raw materials to expansion in manufacturing, services and tourism during
"

I'~ ~

J,

his reigns (Central Intelligence Agency (CIA, 2010).

The fifth prime minister, Tun Abdullah Ahmad Badawi (2003-2009) tried to
move the economy up the value chain by trying to attract investor in high technology
industries and also in pharmaceuticals. The current Prime Minister, Najib Razak
(2009-present) then continues the hard work of Tun Abdullah Ahmad Badawi. He
tried to boost the domestic demand and lower the dependencies toward export.
However, export remains significant especially on oil and gas. New Economic

,: I
i,l.

Model (NEM) was also launched in 2010 to encourage more entrepreneurs to do

business in Malaysia (CIA, 2010).

Malaysia is adopting constitutional monarchy which nominally headed by the

Yang di-Pertuan Agong or also could be regarded as the king. Each sultan among the
nine peninsular states would take turns to be the king and each elected king has 5
year term. The king is also the leader of the Islamic faith in Malaysia. On the other
hand, the executive power is vested in the cabinet that normally led by the prime
minister.

Apart from that, the legislative power in Malaysia is divided into two which
are the federal and state legislatures. In addition, Malaysia legal system is based on
English common law. Federal court reviews decision made by court of appeal.

r:' '

Besides that, Peninsular Malaysia and East Malaysia each has their own high court.

,:;

In Malaysia, the federal government has authority in all matters for example external
affairs, federal citizenship, defence, finance, internal security, commerce and others
except for civil law cases among Malays or other Muslims which is under the
Islamic Law (U.S. Department of State, 2010).

1.2.2 Geography

Malaysia is located in South-eastern Asia where the peninsula is bordering

Thailand while one-third of the island in Borneo is bordering Indonesia, Brunei, the
South China Sea and the South of Vietnam (CIA, 2010). This detail is shown in the
map below.

';'!

'~

Figure 1: Map of Malaysia

Source: Central intelligence Agency (CIA), 2010

The map above shows that the area in a lighter colour of brown is Malaysia.
Malaysia consists of two parts which are Peninsular Malaysia and East Malaysia.
East Malaysia consists of Sabah and Sarawak while the other states in Malaysia are
in Peninsular Malaysia. The two parts of Malaysia is divided by South China Sea.

1.2.3 Economy

Malaysia is a high middle income country that had encounter transformation

from just being a producer of raw material into an emerging multi-sector economy
since 1970. The contri but ion of the other four former prime minister of Malaysia was
already stated previously where it had clearly explained that Tun Dr. Mahathir bin
Mohamad (1981-2003) had contributed highly to Malaysia economy.

In the recent years, the fifth prime minister, Tun Abdullah Ahmad Badawi
(2003-2009)

trie~

to move the economy up the value chain by trying to attract

investor in high technology industries and also in pharmaceuticals. The hard work of
the former Prime Minister was then continued by the current Prime Minister, Najib
Razak (2009-present). He tried to boost the domestic demand and lower the
dependencies toward export. However, export remains significant especially on oil
and gas. In order to encourage more entrepreneurs to do business in Malaysia, the
New Economic Model (NEM) was launched in 2010 where its main motive is to
attract foreign direct investment then under NEM there is the Tenth Malaysia Plan
which outlines new reforms (CIA, 2010). Figure 2 below show the trend in the Gross
Domestic Product (GOP) of Malaysia:

Figure 2: Trend of GDP in Malaysia from 1980-2009

---- 800000.000
700000.000
600000.000
c

500000.000

::!: 400000.000
a:
Q.

300000.000

200000.000
100000.000
0.000

ii'

Source: International Financial Statistic (IFS), IMF, Various Issues

In overall, the Gross Domestic Product (GDP) of Malaysia had increased in

value from RM 53308 million in 1980 until RM 679687 million in 2009. From
Figure 2, The GDP of Malaysia increases only about 8.1 % from 1980 to 1981 and
also only about 8.6% from 1981 to 1982 compare to the increase of 12.6% from
1982 to 1983 and 12.9% from 1983 to 1984 which clearly show that from 1980 to
1982 the growth of GDP was lower. This is because Malaysia experienced high
prices in 1980 and 1981 that were due to external factors. Oil prices increase from 47
percent during 1979 to 66 percent in 1981 and simultaneously the prices of industrial
raw materials also increased rapidly. The increase price of oil by Organization of the
Petroleum Exporting Countries (OPEC) causes powerful pressure on the consumer
prices that was only affected Malaysia in the latter part of the years (Cheng and Tan,
I"

2002).

""

'

')

;; :1'

The GDP of Malaysia slumped in 1985 where it decreased about 2.6%

compared to 1984 and decreased further about 7.6% in 1985. This is mainly because
of the international economic recession during the early 1980s. Because of the
moderate increase in demand and also the tight liquidity position, the capacity of
plants and labour forces are not utilize and as a result prices in 1985 increased at a
slower rate. Inflation rate in Malaysia decelerates and Consumer Price Index (CPI) is
less than 1 percent from 1985 to 1987. This marks a weaker demand condition and
also as a result of the world economic recession, exports and private sector income
depressed as a whole (Cheng and Tan, 2002).

However in 1990s Malaysia economy recovered and GDP began to grow

rapidly where it increased from RM 119081 million in 1990 to RM 300764 million
10

in 1999 where in total GDP increased by 152.6% from 1990 to 1999 which is better
than the growth in the 1980s. The main reason that contributes to the rapid increase
of GDP during this period is the expanding of industrial sector especially in the
manufacturing and services sector (Encyclopedia of the Nations, 2010). Besides that,
New Development Policy (NDP) was also introduced in 1991 which emphasized that
the government would only help Bumiputera with potential and commitment. Other
than that, there was also heavy expenditure on infrastructure as an example the
building of the Twin Tower. The volume of manufactured exports especially
electronic goods and components also increase rapidly (Drabble, 2010).

However in 1997 to 1998, the GDP of Malaysia only has a slight growth which
only about 0.5% as a result of the Asian Financial Crisis that originated in heavy
international currency speCUlation that leads to major slumps on the exchange rate.
This crisis begin with Thai Bath in 1997 and in the end it began to spread rapidly
throughout East and Southeast Asia and affecting the banking and finance sectors.
This event causes a heavy outflow of foreign capital and to counter this particular
situation, Malaysia government pegged its Ringgit at RM3.80 to the US dollar.
Because of this event, every country tried to spend as less as possible thus resulting
in the decrease of export in Malaysia (Drabble, 2010).

After the slight increase in GDP because of the Asian financial crisis, the
policy taken by Malaysian government manage to bring Malaysia out of the financial
crisis and GDP begin to grow rapidly again but later from 2000 to 2001, GDP again
begin to drop about 1.1% in 2001 compare to 2001. The decrease in GDP from 2000
to 2001 is as a result of the global economic downturn and the slump in information
11

technology (IT) sector. As Malaysia growth was almost driven by exports especially
electronics and because of the global economic downturn, Malaysia experienced a
contraction in exports which eventually lead to the decrease in GDP (Malaysia
Canada Business Council, n.d.).

After that, GDP of Malaysia again begins to increase rapidly until 2009 where
there is a decrease in GDP about 7.9% compared to 2008. This decrease is due to the
unexpected drop of the service sector and the drop of the service sector is primarily
influence by the sub-sectors linked to the manufacturing sector. During the first
quarter of 2009, there is a decline in electricity, gas, water and transports and storage
services. All other sectors also show a decline except for construction (Amanah
Mutual Berhad (AMB), 2009).

1.3

....

"

Motivation of Study

The annual growth in the GDP of Malaysia is increasing rapidly especially in

the 1990s but suddenly in 2009 the growth in Malaysia GDP suddenly decreases
;,'
;'~

which is rather astonishing because the GDP during the financial crisis although just
increases it value by a small percentage but it did not decrease as in 2009. Thus, this

study attempts to identify the factors that influence the GDP and the dynamic

relationship between these factors and the country GDP.

Towards the end after identifying the appropriate factors, this study then also
attempts to forecast the GDP in the context of Malaysia in order to analyze the trend
of GDP in the future years. Forecasting is an important tool for the economy today.

12

,.,

",

::.

Forecasting is useful for evaluating the future trend of the economy. Furthennore,
forecast GDP can help to anchor the expectation of the finns and households.

According to Weber (2009), forecasting play an important role in introducing

monetary policy and of course differences in strategy mean forecast also play a
different role. Thus, the growing importance of forecasting to the government
stimulates the passion to build a model that could accurately predict the future
movement of GDP.

1.4

Problem Statement

According to Weber (2009) in a conference organised by Deutsche

Bundesbank, Freie Universitat Berlin and the Viessmann European Research Centre,
he states that the reason central bank have strong interest in forecasting is because of
" '

the substantial and variable lags in monetary policy transmission mechanism as

central banks could not influence current inflation and output. Thus, monetary policy
should be more forward-looking and take a medium-tenn perspective. As a result
from this, forecasts for inflation, output and other macroeconomic variables are
essential input in the monetary policy decision-making process.

The interest in, and demand for, macroeconomic analyses at high frequencies
especially forecast has increases in the recent years including Malaysia as a lot of
commercial banks forecast the country GDP in order to have an idea to Malaysia
future outlook to enable them to charge the appropriate amount of interest and also
introduce new policy that related to banking.

13

,".,

However, conducting analyses and forecasting is not an easy task as different

forecasting models yield different results. Caraiani (2008) forecast the Romanian
GDP using a small Dynamic Stochastic General Equilibrium (DSGE) model while
Angelini, Banbura & Runstler (2008) estimate and forecast monthly national
accounts for Euro Area using a dynamic factor model. Moreover, Debenedictis,
(1997) does a study in British Columbia by constructing a small autoregressive
(VAR) model.

Kogid, Mulok, & Lim (2010) used consumption expenditure, exchange rate
and foreign direct investment as the determinant of economy growth in Malaysia. On
the other hand, Anaman (2004) used government size measured as a ratio of total
government expenditures, total investment and annual growth of labour as the
:"

determinant in Brunei. Kogid, Mulok, & Lim (2010) found that government
expenditure have a role in economic growth but only as a catalyst while Anaman
(2004) found that government size measured as a ratio of total government
expenditures could highly impact the economic growth depending on the size.
Moreover, according to Keynes (1936), government expenditure is one of the
determinants of income.

Besides, Kogid, Mulok, & Lim (2010) also found that similar to government
expenditure, exchange rate also did not play an important role as a determinant for
Malaysia but according to Chong & Tan (2008) the role of exchange rate is still
prevailing especially in the long run for small and open economies. Other than that,
there also exist an enormous theoretical literature on the temporal behaviour of
income and output spanning such areas in public finance, monetary economies,
14

international economies and development economies (Ansari, 2002). In other words,

different countries have different determinants of GDP. Thus, this study in the end
attempt to solve:

1.
11.

What are the determinants that influence the GDP of Malaysia?

Which forecasting models can provide a satisfactory accuracy for the GDP of
Malaysia?

1.5 Objective of Study

1.5.1 General Objective

The aim of this study is to find the determinants of GDP of Malaysia and use
them to forecast the GDP of Malaysia.

1.5.2 Specific Objective

The specific objectives of the study are:

" ?

1"

.,

."

I,'

1.

11.

HI.

To examine the determinants ofGDP in Malaysia.

To forecast the GDP of Malaysia.

To evaluate the forecasting accuracy of the forecasts.

1.6 Significance of Study

Forecasting the GDP of Malaysia could generate future values on GDP and
these futures values could be used to determine the trend and behaviour of the GDP.
The benefit of being able to determine the trend ofGDP is the government would be
able to know the condition of the GDP for the following quarters or even years. This
15

\'

.1.

advantage enables the government to find policies that match the condition of the
economy for the following quarters or years. Thus, forecasting the GDP is important
as this could be a tool to help the government to evaluate the future economy and
find out whether the new policy and new approaches that was about to be introduce
suits the economy.

Besides that, accurately forecast the future trend of the GDP also could alert
the government if there is going to be a slump in the country GDP. Then, the
government could implement plan to counter the slumps earlier to avoid the country
suffering from the economic downturn. On the other hand, the future trend of the
economy also can help the central bank introduce proper monetary policy especially
when there is a decreasing trend detected in the future value of the GDP.

Indirectly, this study is also important to the public as when government and
the central bank take the appropriate measures according to the future trend of the
GDP, the public will be benefited as the government would always find the most
beneficial method that would lead the country and the people better off. Thus, the
public would be spare from any economy slow down and would be able to prosper
when the economy is booming.

In overall, forecasting the GDP is important for the government especially

when new policies is intended to be introduced and to prepare for any economic
downturn in the near future.

16

"

}'

1.7

Structure of the study

This paper is organized into five chapters in which Chapter One will
briefly discuss on the background of Malaysia then later throughout Chapter One, it
will deals with problem statement, the objectives, motivation of study, significance
of study and also the scope of study. Then, Chapter two contain literature review of
the theoretical studies of forecasting GDP and also review on previous empirical
studies. Then Chapter three will describe on the research methodology used in the
study. Later, Chapter four explains and report the result of the empirical analysis and
lastly, Chapter Five concludes the study conducted along with some policy
recommendation.
I
I

I:' I',
"
~,

:;

II

.I'

"

",

17

,.

Chapter Two
Literature Reviews
2.0

Introduction

Gross domestic product (GDP) is the basic measurement of a country

economy's performance. GDP is a measure of the value of all the goods and services
newly produced in a country during some period of time (Taylor, 2007). GDP can be
defined in three ways but in concept it has the same meaning. First, GDP is equal to
the total expenditures of all final goods in a given time. Second, GDP is the sum of
all the stages of production value added, by all the industries within a country. Third,
GDP is the sum of the income generated by the production in the country in the
period.

Time series forecasts are used in all kinds of economic activities which include
setting of monetary and fiscal policies, state and local budgeting, financial
management and financial engineering. The key element that must present in
economic forecasting is selecting the forecasting model and also assessing and
communicating with the uncertainty associated with a forecast, and guarding against
model instability. One of the types of economic forecast is forecasting GDP and its
determinants. The determinants used to forecast GDP is mostly national account data
which include export, import, real effective exchange rate and more. Besides that,
industrial production index is also used as a determinant basically because industries
are one of the sectors that contribute to a country GDP. However, the significance of
the determinant towards GDP varies according to country, state and region. Some
18

I
I
1:1

'

I'

studies that was done also show that there are cases when money could also affect
the GDP or in other words money is also used as a determinant of GDPJ.

In order to understand more on the result found by different studies, this

literature review is divided into 5 sections which consists of 2.0 which is the
introduction on generally what does GDP and forecasting means. While, 2.1 is the
theoretical framework on how the determinants of GDP would lead to the changes in
GDP, 2.2 encompasses the empirical testing procedure that would be separated into
two parts which the first part would discuss on the empirical model used to forecast
the GDP and the second part is on the method used to forecast. Then, 2.3 consist of
the empirical evidence or finding of previous study lastly 2.4 is the concluding
remarks.
(

2.1

Theoretical framework
,I" I .

1i ,;:,,,

According to the previous studies, some of the determinants of economy

J ::

t ,,,
~ ! .:~

~~

growth mentioned are:

"

I
I
j

2.1.1 Stock Market

Stock market return is one of the determinants of Gross Domestic Product

(GDP). Stock market return is often represented by stock price or share price. Stock
price is negatively correlated with interest rate. Changes in interest rate could affect
stock price through substitution effect. This is mainly because that higher interest

See Pigou (1943) and Patinkin (1965) which the relation of real balance effect originated.

I
I

19

il

;1

rate would result in contractionary monetary policy that would cause the return of
stock to become lower. As interest rate goes higher fixed income securities became
more attractive than holding stock in other words people would prefer to save their
money to earn a higher return than to loan money and suffer from the high interest
when they need to pay back the loan (BaneIjee & Adhikary, 2009). As loan
decreases, economic activities that involve transaction and investment will
eventually decrease and in the end would lower the output of the country. The lower
the stock price indicates the lower the investment as investment is lower than the
output is also affected as investment contribute to the country GDP. This conclude
that stock market and GDP is positively correlated.

2.1.2 Real Activity

Another determinant of GDP is changes in real activity which is often

represented by industrial production. Changes in real activity is related to interest
rate as a decrease in interest rate would cause an increase in investment and therefore
in the future production. As discuss above, interest rate could also affect stock price
now through variation in future product (Peiro, 1996). The relationship between
stock price and industrial production had been proven by Fama (1981, 1990) that
shows the close relation between real returns and growth rate in industrial production
empirically using annual data. Indirectly, changes in industrial production would
affect the stock prices and also the interest rate that eventually would lead to the
changes of the country output. Thus, as real activity in the country increases, the out
or GDP will eventually increase. This shows a positive relationship between real
activity and GDP.
20

.,"

2.1.3 Money Supply

Money supply is positively correlated with GDP which is illustrated by Arnold

(2008). He state that a change in the aggregate demand and thereby change the price
level and also GDP in the short run. This is to say that an increase in the money
supply would shift the demand curve to the right which will move the economy to a
higher point. On the other hand, if money supply is decrease then lower level of
GDP will be produce

2.1.4 Exchange Rate

Apart from that, exchange rate is another GDP determinant. According to Yau
and Nieh (2009), there are two theories that are about the relationship between
exchange rates and stock prices which is the traditional and portfolio approach.
Traditional approach state that depreciation in the domestic currency would cause
the local firm to become more competitive that lead to an increase in their exports
and would yield higher stock price. This approach implies that exchange rate and
""

stock price is positively related. Meanwhile, the portfolio approach states that an
increase in the stock prices would induce the investor to demand more foreign assets
which in the end could cause an appreciation in the domestic currency. This
approach on the other hand implies that exchange rate and stock price are negatively
related.

Empirically, there are a number of researchers that proven that a significant

relationship exist between the exchange rate and stock prices. Mok (1993) found a
weak bi-directional causality between stock prices and the exchange rate. Nieh and
21

",

Lee (2001) also found bi-directional causality between stock prices and exchange
rates but only in the short run. Besides them, there are a few other researchers that
either found a weak association or a zero association between stock prices and
exchange rate (Franck and Young, 1972; Bartov and Bodnar, 1994; Fernandez,
2006). The changes in exchange rate would eventually affect the stock prices. Stock
prices on the other hand could influence the changes in GDP as stock price reflect
investment. The lower the stock price indicates the lower the investment as
investment is lower than the output is also affected as investment contribute to the
country GDP.

On the other hand, overvalued exchange rates are associated with shortages of
foreign currency thus will damage the economic growth. In other words, an increase
in the undervaluation will boost economic growth as well as a decrease in
overvaluation. This indicates that exchange rate and economic growth are negatively
correlated.

2.1.5 Interest Rate

Interest rate is also one of the most important determinants in GDP. As had
shown from all the above determinants, almost all of the determinants have the
present of interest rate as a transition mechanism. As interest rate increases, holding
stock would become unattractive as they cannot gain much as a result of the high
interest rate which causes them to have to pay a lot if they loan a lot. This
phenomenon induced people to save rather than invest. Saving generally is a kind of
withdrawal which could decrease the amount of output produce. The relationship
between interest rate and GDP was portrayed by Obamuyi (2009) that imply the
22

behaviour of interest rate is important in economy growth that normally represented

by GDP or GDP growth as interest rate could affect investment and investment could
affect the output of the country.

2.1.6 Trade

In addition, trade is also a determinant of GDP and this sense trade constitute
of export and import. According to the Keynes (1936), export and import is a factor
that could influence the amount of the income of the country. This is because when
the demand of export for other foreign countries increases thus directly it could
increase the income of the country whereas indirectly is through the changes of the
term of trade. Exports also would likely to alleviate foreign exchange constraint thus
provide greater access to international market. In other words, as more goods and

,',

services were bought by other countries, the goods and services would be paid by
them thus when they pay for its then the income of our country would eventually
increase. Other study that emphasizes the importance of exports towards the country
output or economy growth is Dritsaki, Dritsaki & Adamopoulos (2004) which found

'.
"
.,

"

that economic growth, trade and FDI appear to be mutually reinforcing. Economy
growth used in their study is GDP. Other studies that support the significance of
export to GDP are Sentsho (2000), Fugazza (2004) and Awokuse (2002).

2.1.7 Consumption Expenditure

According to Kogid, Mulok & Lim (2010), consumption expenditure is one of

the variables that play an important role as a determinant factor to a country
economic growth in this sense GDP in Malaysia. Besides that, Keynes (1936) also
23

I'

discovered that consumption expenditure is one of the determinants for the income
of a country. As the consumption expenditure increases, this means that more goods
and services are being consume and injecting more income or revenue to the market
thus increasing the revenue of the country. There are also a number of studies done
regarding the relationship between consumption expenditure and economy growth
such as FoIster and Henrekson (1999) and Kweka and Morrissey (1998) although
both of these studies reported no evidence of relationship between economy growth
and consumption expenditure for their respective country.

2.2

Empirical Testing Procedures

2.2.1 Specification of models

Some methods that were used by previous studies to model the determinant of
GDP are discussed by Anaman (2004) that assume a Cobb-Douglas functional form
and restate the economy wide production function as:

Yt

= Boexp (B 2+ 3
B + B3
4
1

G4
)

(TEXPORT)B4 (TLABOR)Bs

(2.1)

where exp denotes the exponential operators, G refers to government size where it is
defined as government expenditure divided by GDP, TEXPORT is the total annual
level of exports, TLABOR is the total annual level of labour inputs, TCAPITAL is the
total annual stock of capital inputs, ASIANFC is the dummy variable with a value of
1 for the years. By taking logarithm in equation (2.1), the empirical model used in
this study is as follows:
24

...

GROWTHt

= Bo + B1 (GOVSIZ)t + Bz (GOVSIZ2)t + B3 (GOVSIZ3)t +

B4 (GTEXPORT)t + Bs(GTLABOR)t + B6 (JNVGDP)t + B7 (ASIANFC)t + Ut

(2.2)

where GROWTH is the annual growth of real gross domestic product, GTEXPORT is
the annual growth rate of the real value of total exports, GTLABOR is the annual
growth rate of total labour force, and GTCAPITAL is the annual growth rate of the
real value of total capital stock. GOVSIZ is the relative size of government defined as
the ratio of government expenditures to gross domestic product. GOVSIZ2 is the
square of GOVSIZ, GOV5)JZ3 is GOVSIZ raised to the third power and INVGDP is
the lagged ratio of total investment to gross domestic product, and U is the error term
that is assumes to be normally distributed.

It is hypothesised that government size will impact the economy growth in a

cubic function. Initially small relative size of government hampers economic growth
while medium-sized government accelerates economic growth through the provision
of basic infrastructure and improved legal framework, and an increased growth of
total exports, labour and investment inputs are hypothesised to lead to increase
economic growth. The growth on human-made capital inputs on the other hand is
expressed in the form of a ratio of investment to GDP. Investment represents change
in total capital stock and should be divided by total capital stock to derive growth
rate of capital. So, small relative size of government constitutes a negative sign while
medium-sized government, exports, labour and investment is positively correlated
with economic growth.

25

I',

Gounder (1999) also uses the same method as Anaman (2004) that study the
effect of military coups on Fiji economic growth for the period 1968 to 1996 which
the variables used in the study are annual growth rate of national income, annual
growth rate of labour force, total investment to output ratio, private investment to
output ratio, government investment to output ratio, annual growth rate of exports,
political instability variable of military coups.

Besides that, Sentsho (2000) uses a similar method as the previous authors to
assess whether export revenues derived from an enclave sector like the case of
mining in Botswana can lead to significant and positive economic growth in a
country. The variables used in the model are real GDP, ratio of gross domestic
investment to GDP and labour force. The unconventional inputs used in this study
include aggregate export, primary export, manufactured (non-traditional) export,

,
I

imports, private sector consumption in real GDP, government sector, previous period
growth in real GOP and world GOP. The sample period is from 1975 to 1997 and to
capture the economic boom that came with the opening of the Jwaneng diamond
mine and the construction of its town a dummy variable is used.

On the other hand, Kogid, Mulok, & Lim (2010) begin the functional exact
relationship between the dependent variable and independent in logarithmic form (L)
where Yt is a function of Xit which can be specified as below:

(2.3)

where

Yt

is LGDP at time t,

Xit

is Log Consumption Expenditure (LCE), Log

Government Expenditure (LGE), Log Export (LX), Log Exchange Rate (LER) and
26

,I'

Log Foreign Direct Investment (LFDI) at time t, i

1, 2, 3, ... , n. Thus, to allow for

the inexact relationship between economic variables, the deterministic economic

growth function is modified as follows:

(2.4)

where

is known as disturbance or error. The disturbance term

may well represent

all those factors that affect economic growth but not taken account explicitly. FDI
contributes largely to the development of East Asian economy and from his
literature, real GDP was found to have a positive impact to FDI inflow. Export on
the other hand is the most research determinant as one of the reason because most
developing countries practice export promotion. Export is positively correlated with
income as exports could increase the income of a country. Expenditure and exchange

I',

rate also constitute a positive relationship. However, according to them theory itself
."
I

is not enough as theory provide little evidence. Besides that, Chen & Feng (1996)
also uses the same model using average growth rate in real GDP per capita, political
variables to capture regime instability, political polarization, and government
repression and control variables to measure economic conditions.

2.2.2 Forecasting Models

There are a few models that mostly used by researchers to forecast the GDP.
Some of these models are shown in the next subsection:

27

2.2.2.1

Vector Autoregressive (V AR) Model

Saraogi (2008) did a research to forecast the quarterly growth rates in the GOP
of Australia using a VAR model, which has some obvious benefits over a pure
simultaneous equation system. In this model, some variables are treated as
endogenous and some as exogenous. This model is based on Sims2 The model as
research by Saraogi (2008) is:

Yt

= a + PHClt _ i + yPClt - i + TJBl t - i + 8ESIt _ i + Ct

(2.5)

where, Yt == Quarterly growth rate in GOP at constant 2000 US$

Intercept tenn
I'

Human Capital Index with i th time period lag

HClt - i

th

PClt - i == Physical Capital Index with i time period lag

Bl t -

i =

ESl t - i

Banking Index with ith time period lag

P, y, TJ, 8
Ct

External Sector Index with ith time period lag

Regression Coefficients

Error tenn

2 Sims ( 1980) states that if there is simultaneity exist among the variables then they should be treated on an equal
footing where there should be no distinction between the endogenous and exogenous variables.

28

"

Reimers and Seitz (2003) assess the predictive content of MI using different
types of VAR models. The first model use is unrestricted VARs as this model is a
good empirical representation of economic time series as long as lags are included3
Their model is:

(2.6)

where Xt is the vector endogenous variable, MI,

the matrix deterministic terms,

especially the intercept term and linear deterministic trend, Al to Ao are the
symmetric coefficient matrices and 0 is the selected lag order. However if the
variables are cointegrated but not stationary, a Vector Error Correction (VEC) is
used. VEC becomes a VAR model in the first differences.

Then from a study by Barhoumi et aL (2008), they use a quarterly type of VAR

i'

'"

'I

model to forecast. They run bivariate VARs including GDP and the quarterly
aggregate of a single monthly indicator and later they average the forecasts across
indicators. Their model is:

QZi,t
- fli

+ '"s=1
,,(,Pi A ZQ
+ ei.t'
Q
s i.t-s

i = 1, ... , k ,

produced. Then Pi represent the lag length.

See Canova (1995) on VAR specification, estimation, testing and forecasting.

29

(2.7)

2.2.2.2

Dynamic Stochastic General Equilibrium (DSGE) model

Caraiani (2008) study the Romanian GDP using a small DSGE modeL The
model consists of a finite number of representative agents characterized by an
infinite life. Each agent will maximize the expected lifetime utility. Consumption,
investments and labour effort will be optimally chosen under the constraints given
by its income. Based on the model studied by Caraiani (2008), the model is given by:

00

maxEo [ Lt=of3

t Ct1-11 -1
1-1')

- ANt

(2.8)

where f3 is the discount factor, Ct is the consumption, 11 is the relative aversion

coefficient, Nt is the number of hours worked and A is the parameter that symbolizes
the utility function.

2.2.2.3

Dynamic Factor Model

The concept of this model is based on the assumption that macroeconomic

variables are better described with small unobserved common factors. Krajewski
(2009) did a study using dynamic factor model that is based from Stock and Watson
(1998) where he let Yt stand for a variable and Xt express the vector of N variables
that contain useful information to forecast Yt. In the model all variables Xit which
include output and sales, construction, domestic and foreign trade, prices and labour
market, budgetary and monetary policy that contain in vector X t may be expressed as
a linear combination of current and lagged unobserved factors fit .

Xu

= AiCL)ft + eit

for i

= 1,..., N,

(2.9)

30

:.

where

It

stands for vector f of unobserved common factors at moment t, and Ai (L )

represent lag polynomials and eit express an idiosyncratic error for variable xu- In
the end, Yt may be noted as the function of current and lagged common factors
contained in vector It and the past values of variable Yt with the following formula:

Yt

= P(L)!t + y(L)Yt + et

(2.10)

Model described in equation (2.10) and (2.11) are both dynamic factor model.

Besides Krajewski (2009), a study on forecasting the GDP of Austrian is also

done by Schneider and Spitzer (2004) using this model with the only difference is
that now the model is generalized. Their model is written as:
,',

(2.11 )

,JI

i;'

where Xit is called the common component and include variables such as national
account data, WIFO quarterly survey, monthly survey data, prices, foreign trade,
labour market, financial variables and industrial production. On the other hand

Cit

is

the idiosyncratic component. In addition, bi(L) is a vector of lag polynomials and

lastly

J1.t

is a q-dimensional vector of common stocks. In their study, the q-

dimensional process is assumed to be mutually orthonormal white noise with unit

variance Cit on the other hand is orthogonal to J1.t-kfor any k and i.

31

,.,

,I>

A research by D'Agostino, McQuinn and O'Brien (2008) was also done by

using the same model and applying it to now cast the GDP of Irish. Their model is
based on the model of Giannone, Reichlin and Small (2007)4. The model is:

(2.12)

(2.13)

(2.14)

where in equation (2.l3),

Xt

include economic activity, price dynamics, business and

consumer sentiments surveys and financial indicators and

orthogonal components, the common component x t

Xt

and the idiosyncratic

component ~t. The common component is the product of a n x

A and T x 1 vector of latent factors

is the sum of two

matrix of loadings

f! . Meanwhile, the idiosyncratic component is a

multivariate white noise with diagonal covariance matrix L{' On the other hand,
factor dynamics are described in equation (2.14) which is a VAR (p).

2.2.2.4

Univariate Autoregressive Integrated Moving Average (ARIMA)

A research by Hoehm, Gruben and Fomby (1984) was done to forecast the
economy of Texas in which one of the models used is an ARIMA model. This model
is selected as it treats each variable in isolation no matter in estimation or in
forecasting. The model that they come out with is:

See Giannone. Reichlin and Small (2007) about nowcasting the real time informational content of
macroeconomic data where they developed a formal method to evaluate the marginal impact that intm-monthly
data releases have on current-quarter forecast of real GDP growth. Their model is use in the study of Agostino.
McQuinn and O'Brien (2008)

32

"

"

'"
,!!

(2.15)

where L is the lag operator. Yt is the natural logarithm of the series, the variables in
the series include Texas Industrial Production, Consumer Price Index, Payroll
Employment, Household Employment, Texas Labour Force, Deflated Personal
Income and Deflated Retail Sales and at is a normally distributed unobservable
random variable with zero mean, finite and constant variance and have zero
autocorrelation at all lags. There are p autoregressive term (lagged y' s) and q
moving average terms (lagged a's).

2.2.3 Empirical Method

Before the variables are estimate, all the variables must be makes sure that they
J'
,I;

are in the right order. Besides that, only cointegrated variables can be used to
forecast as this means that they exhibit long run relationship.

2.2.3.1

Stationary test

"
d

According to Lu (2009), a time series that is about to be analyzed should be

make sure that it is stationary before specifying a model. Lu (2009) had used KPSS
and ADF test to investigate whether the time series are stationary. ADF test is used
to verify whether the series is stationary while KPSS is to measure whether unit root
exist.

33

Debenedictis (1997) also tested for unit roots. Like Lu (2009), she also adopt
ADF test but if the errors are not independent and seasonal data is detected then the
basic ADF 5 test needs to be modified6

Besides that, Anaman (2004) like the two other researchers also conduct unit
root testing. He adopts DF or ADF, Phillips-Perron (PP), Kwiatkowski-Phillips
Schmidt-Shin (KPSS) and Ng-Perron (NP) test.

2.2.3.2

Johansen Multivariate Cointegration Test

Anaman (2004) apply Johansen Multivariate Test to test the movement of the
variables in the long run. They only use the variables of the same integration level to
test the presence of cointegration level. The lag structure is determined automatically
I',

by the statistical package. The core movement in the long run are as Johansen (1988)

"
1

"

;1'
I

and Johansen and Juselius (1990) and can be defined as follow:

(2.16)

Where,

Mi = -(1-

TIl - ... - TID, (i =

1, ... , k - 1),

And

II = -(1-

5
6

TIl - ... - TID

The basic Augmented Dickey-Fuller test that was presented by Dickey and Fuller (1979, 1981)
A modified ADF test is based on Said and Dickey (1984) research.

34

Model (1) is expressed as a traditional first difference of a V AR-model except

for llXt - b thus coefficient matrix

is used to be investigated to find out whether it

contains information about long-run relationship among the variables in the data.

The rank of

n depends on two likelihood ratios which include Trace test and

Maximum Eigenvalue test. Trace test as expressed as:

(2.17)

where T denotes the number of valid observations for estimation use and ~ is the ilb
largest estimated eigenvalue.

On the other hand, Maximum Eigenvalue is expressed as:

(2.18)

where T denotes the number of valid observations for estimation use and 1,.+1 is the
largest eigenvalue at r

+ 1.

After the variables in the equation are prove to be stationary and cointegrated
with each other then the next step is to forecast the mode1.

2.2.4 Forecast Criteria

Evaluating the forecast accuracy is also important as it will give an idea on

how well the model could represent the real economy. Thus, a few test to determine
the accuracy of the model is conducted.

35

2.2.4.1

Information Criteria

On the other hand, according to Krajewski (2009), in practice usually the

number of factors necessary to represent the correlation among the variables is
usually unknown. Krajewski (2009) determine the number of factors empirically
using the information criteria suggested by Bai and Ng (2002) which is given as:

ICi (k) = In (V(k)) = k ((N;;) In (::T))

IC2 (k) = in (V(k)) = k ((N;TK) in C~T) ,

IC2 (k) = in (V(k)) = k C:~;T)

(2.19)

(2.20)

(2.21)

NT =

where V(k) is the residual sum of squares from k-factors model and C

min{vNv'T}. N which the number of factors yielded from the equation or test will

represent the correlation among the variables.

2.2.4.2

Root Mean Square Error (RMSE)

In the study of Guegan and Rakotomarolahy (2010), using five years of

vintage data, they adopt RMSE to evaluate the accuracy ofthe forecast by computing
the RMSE for quarterly GDP flash estimates. The RMSE criterion used for their
study on the final GDP is:

(2.22)

36

Where T is the number of quarters between the period7 that their estimated and Yt is
the Euro area flash estimate for quarter t. Lower value of RMSE indicate a better
model.

In the study done by Andersson (2007), he also uses RMSE as one of his
method to measure the accuracy of the forecast. According to Andersson (2007),
RMSE is the most frequently used measure and is known to be more sensitive to
outliers than MAE8. However RMSE is the only one he uses to rank the performance
of the model. After that, F test is use to check whether the differences in the forecast
ofRMSE is significant. The F-test is formulated as:

F -

"H

""i=l eli

(2.23)

"H
z'
""i=Z eZi

The larger value of the forecast RMSE is put in the numerator. The null hypothesis is

"

for equal forecasting performance for the two models being compared. The intuition
is that the F-value will equal unity if the forecast RMSE from the two models are
,i'

equal, while a very large F-value implies that the forecast RMSE from the first
model is substantially larger than the forecast RMSE from the second model.

2.2.4.3

Mean Absolute Percentage Error (MAPE)

Gupta (2006) also adopted MAPE in order to evaluate the forecast accuracy.
The statistic of the MAPE test adopts by Gupta (2006) can be defined as:

Time period define in their study is QI 2003 and Q2 2007, T=18.

It is the averages ofthe absolute values ofthe out-of-sample forecast error.

37

( 2:.NLabs (At+n-tFt+n))
At+n

100

(2.24)

where abs stands for the absolute value. For n

to 2005:4; and for n

1, the summation runs from 2001: 1

2, the same covers the period of2001:2 to 2005:4 and so on.

Note that, At+n denotes the actual value of a specific variable in period t

+n

and

is the forecast made in period t for t+n. Percentage errors are not scale-

t Ft+n

independent thus they are usually used to compare performance across different data
sets. The higher percentage error indicates lower accuracy.

2.2.4.4

Encompassing tests

This is another method to access the relative performance of the models which
was study by Barhoumi et al. (2008). In this test, two alternative models 1 and 2 are
based on regression of the actual data Yt on foreca')t h~ and fz~t from two models.
Q

The equation is:

(2.25)

and

where It gives the optimal weight of model 1 in the combined forecast. However, in
the extreme case, a value of It

= 1 indicates that modell dominates model 2.

38

I',
I,

2.3

Empirical Evidence

Anaman (2004) conduct a study in Brunei Darussalam to find the country

determinant using yearly data from 1974 to 2002 obtain from the various issue of
Brunei statistical yearbook, 1974 to 2000 for private sector labour force and 1971 to
2000 for government workforce. Multiply regression analysis based on relatively
new cointegration technique was used to construct a variant of the neoclassical
model. The results from his research show that, the growth of export significantly
influenced long-run economic growth rates and the other main factor is the relative
size of government where it influence the long run growth rate in the form of cubic
function. Large government sizes impeded economic growth while moderate
government sizes enhanced economic growth.
, l",

Kogid, Mulok, Lim and Mansur (201 0) conduct a study to determine the
economic growth factors in Malaysia using yearly data from 1970 to 2007. They use
cointegration analysis and causality approach and also ECM to analyze the
relationship between economic growth and the determinant factors. They find that
consumption expenditure, government expenditure, export, exchange rate, and
foreign direct investment cause economic growth in the short run. However,
individual test conducted point out that only consumption expenditure and export
cause economic growth. Thus, they conclude that consumption expenditure and
export play important role as determinant factors to economic growth.

Gupta (2006) conduct a study in South Africa to forecast the country economy
using quarterly data from the period of 1970: 1 to 2004:4 with an out of cast sample
from 2001:1 to 2005:4 with a Bayesian Vector Error Correction Model (BVECM)
39

and find that BVECM in general except for long term interest rate measure, produces
the most accurate forecasts relative to the alternative models so it is best suited for
forecasting South African GDP. The Mean Absolute Percentage Error (MAPE)
values from the VECM are less than those generated from VAR model. The VECM
and BVECM correctly predict the direction of change for the entire variable.

This finding is later supported by Uu, Gupta & Schaling (2010) that also
forecast South Africa economy using quarterly data from the period of 1970: 1 to
2000:4 and with an out of sample forecast from 2001:1 to 2005:4. They estimate the
model with the maximum likelihood technique on real gross national product (GNP)
then based on a recursive estimation using the Kalman filter algorithm from the
model are then compared with the out of sample forecast from classical and
Bayesian variants of the VAR. They say that in general, the estimated hybrid DSGE
model outperforms the classical VAR, but not the Bayesian V ARs. The Root Mean
Square Error (RMSE) generated from the BVAR is much smaller from both the
hybrid model and unrestricted V AR.

However in the earlier studies, Debenedictis (1997) does a study in British

Columbia where a small autoregressive (V AR) model is construct using quarterly
data from the period of 1961: 1 to 1991:4 and an out of cast sample from 1992: 1 to
1995:4 and find that V AR yields the least biased forecasts for employment and
British Columbia GDP while ARlMA yields the least biased estimates for Canada
GDP and prices. The result indicates that 98% and 89% of the actual data fell within
the 95% confidence intervals for ARIMA and VAR models respectively. Thus, VAR
is a promising forecasting tool in the case of British Columbia.
40

Schneider & Spitzer (2004) foreca')t the GDP of Austrian using a mixture of
quarterly and monthly data from the period of 1988: 1 to 2003 :2. They use a
generalized dynamic factor model to produce the short term forecast and the model
follows the frequency domain approach. They state that broad movements of the
business cycle are predicted correctly as the results from Pesaran-Timmermann test
show that the change in the GDP growth rates direction is correctly predicted. On the
other hand, the forecasting error for two steps ahead forecast is smaller than one step
forecast. In conclusion, Factor model perform better in small data set, the number of
dynamic factor could impacts heavily on forecast performance.

In addition, from the result obtain by Caraiani (2008) where he forecast the
Romanian GDP using a small DSGE model using quarterly data from 2006: 1 to
I',
I,

2007:2. The model is forecast based on the posterior distribution of the model

r
'1. 1

parameters as a result from the Bayesian estimation. He suggests that DSGE model
can have a good performance for the trend of the growth however its overall

...

performance depends on the level of temporary productivity shocks. The projection

,I'

did by him show that a growth rate of almost 6% which is in line with other
estimates. Thus, DSGE model also could be used to forecast the GDP of Romania.

Angelini, Banbura & Runstler (2008) estimate and forecast monthly national
accounts for Euro Area using quarterly data from 2000: 1 to 2006:2 and also using
monthly data for a different kind of method from 1998: 1 to 2006:6. The growth in
euro area monthly GDP and its components was estimate and forecast using a
dynamic factor model then the model is extend to integrate interpolation and
forecasting together with cross-equation accounting identities. They find that
41

forecasts and monthly estimates are consistent which will be advantageous for
monitoring economic development in real time. For GOP and a number of
components, the model beats forecasts from time series model based on quarterly
data and from forecast average from bridge equation. As regards to GOP, AR (1),
V AR and bridge equation improved upon naive forecast where the gains in RMSE
are close to 20%.

Mittnik & Zadrozny (2004) forecast the quarterly GOP at monthly intervals
using monthly IFO business condition data in Germany. They use quarterly data
from 1970: 1 to 2003:4 and monthly data from 1970: 1 to 2003: 12 to evaluate the
GOP forecast. They use Kalman filtering method that based on maximum likelihood
estimation method to forecast. They state that monthly GOP forecast are feasible and
produce better short term GOP forecast. Meanwhile, Quarterly data produce better
GOP forecast and aggregation affect the forecast. IFO variables improve long term
GOP forecast and to estimate larger models we must extent to mixed-frequency
forecasting. Choosing 0 options and unit option in analogous test improve accuracy
of GOP forecast.

On the other hand, Barhoumi et al. (2008) predict the growth rate of quarterly
French GOP from 1988:3 to 2006:4. The model is designed to integrate the monthly
economic information through bridge models for both demand and supply sides of
GOP. However different variables have different starting period depending on the
availability of the data. They find from the result they obtain from the euro area
countries that exploiting the timely monthly release fare better than quarterly models.
Besides that, a finding by Golinelli and Parigi (2008) suggest that the relevance of
42

indicators tends to varnish at longer forecasting horizons however, the forecast ofthe
first GDP release obtained from real-time data and from the latest available data do
not differ significantly.

Hoehn, Gruben & Fomby (1984) forecast the economy of Texas with some
time series method using quarterly data from 1969 to 1980 to forecast and from
1981 : 1 to 1983:2 evaluate the performance using a VAR model and they find that
mean errors for all variables except labour force generate by two alternative ARIMA
models were negative. Besides that, RMSE also generally exceeded the standard
error of the within sample often quite substantially. Among the seven series studies,
they state that the two employment series seem the most important for the regional
forecaster to watch.

2.4

Concluding Remarks

As a conclusion, chapter has reviewed literatures that are taken from previous
study which is related to forecasting and now casting the GOP. From previous
researches, it is known that there is a lot of model used to predict the GDP of a
country. These models consist ofVAR, BVAR, Dynamic Factor model, DSGE and
ARlMA models. Among the model reviewed Bayesian VAR model perform better
than the others.

In addition, it is mention by Lu (2009) that the data need to be stationary

before proceed to any kind of estimation. Most importantly, every forecast the
accuracy must be determine whether to compare it with real time data or forecast

43

estimated from unrestricted model. In the end, according to Andersson (2007) RMSE
is a better tool to measure forecast accuracy as it is more sensitive to forecast outlier.

So, the research is intended to forecast the GDP of Malaysia using the VAR
model. In this study in terms of the methodology, unit root test will be conducted as
to determine whether the data is stationary. Then, Johansen Co integration test will
implemented as to find out whether the determinants used is cointegrated as they
need to be cointegrated in order for them to have a long run relationship.

Lastly, RMSE test will be used to evaluate the forecast accuracy as to find out
how much thus it deviates from the original values.

I',

44

Table 2.1: Summary of Literature Review

Author
Angelini,
Banbura&
Runstler,
(2008)

Data
Countly
Euro Area

Variable
Euro area Data set
consisting Euro
area economic
activity
survey data
Euro area trade
values
financial data
International
economy

Sample Period
Evaluated
over the
period
2000: 1
2006:2

_Type ofData
Quarterly data

Recursive
estimation
start 1993: I

1998:1
2006:6

45

"- z

-::-__ -:: :: :'.

Monthly data

Methodology
Use naive
forecasts and
first-order
autoregressive
processes.
Vector
Autoregressions
(V ARs)
Bridge equation
DFM considering
both with and
without
constraint

Chow-Lin
method to a
single equation.

Findings
For GOP and a
number of
components, the
model beats
forecasts from time
series model based
on quarterly data
and from forecast
average from
bridge equation.
As regards to
GOP, AR (1),
VAR and bridge
equation improved
upon naive forecast
where the gains in
RMSE are close to
20%.

'ft

Table 2.1: Summary of Literature Revie,!(c::ontinued)

Data
Author
Gupta,
(2006)

Variable

GOP
consumption
investment
short and long term
interest rates

Country
South Africa

Salllple Period
1970: 12000:4

Type ofOata
Quarterly data

Methodology
Estimate using
a BVAR and
BVECM

CPI.
2001 :12005:4

46

'::""'..---=-: :;

:::::~---:.-

':;;

Eight quarter
ahead forecast
to compare the
accuracy of the
forecast relative
to the forecast
generated by an
unrestricted
VARanda
VECM.

Fjllding
BVECM in general
except for long
term interest rate
measure, produces
the most accurate
forecasts relative to
the alternative
models so it is best
suited for
forecasting South
African GOP.
The Mean
Absolute
Percentage Error
(MAPE) values
from the VECM
are less than those
generated from
VARmodel.
The VECM and
BVECM correctly
predict the
direction of change
for all the variables
over the period of
2004: 1 and 2005:4.

L". '-

15('

'

1'5'

'Q

Table 2.1: Summary of Literature Review (Continued)

Data
Author
Andersson,

Variable
Swedish real GDP

Coun~

Swedish

(2007)

_SaI!lple Period ..
1993:12006:4

47

- "-

--

. Ii'pe.<>f[)a,ta
Quarterly Data

Methodology
The forecasts
from the
different
models are
generated using
an expanding
information
window
approach.
The different
models are
evaluated using
standard
forecast
evaluation
criteria.

Finding
The vector
autoregressive model
with 1 lag and
confidence in the
manufacturing
industry as
explanatory variable
perform best for
forecast horizon t+ 1.
2 lags best for
forecast horizon t+4
and 3 lags are best
for forecasting
horizons t+8 and
t+12.
However, when the
differences in
forecast RMSE
between the best
performing model
and the second best
are tested for
significance, the best
performing models
are not significantly
better than the
second best.

tTf

Table 2.1: Summary of .Literature Review (Continued)

Author
Liu, Gupta &
Schaling,
(2010)

Variable
Real GNP
Consumption,
investment
Hours worked.

Data
Country
South Africa

Sanlple Period
1970: 1
2000:4

2001:1
2005:4

48

':':'

':-,-~::

:::-:-~":.~

-. -;- -.':-.

of Data
Quarterly data

M~hodology

Recursive
estimation using
Kalman filter
algoritm

Out-of-sample
forecasts from
the hybrid model
compared with
the forecast
generated from
the Classical and
Bayesian variants
of the VAR

Findings
In general, the
estimated
hybrid DSGE
model
outperforms the
classical VAR,
but not the
Bayesian
VARs.
The Root Mean
Square Error
(RMSE)
generated from
the BVARare
much smaller
from both the
hybrid model
and unrestricted
VAR.

.........-.-

---~-----

"'-'~'---"--.--------'-~'-

~;;.';=;;';-;'

~"":":'":;'~.':"~- -~--."

... Woo...

"..

Sample Period
1961: 11991:4

Type of [)ata
Quarterly data

Quarterly Data

"

,.

Table 2.1: Summary of Literature Review (O.mtinu(ld)

Data
Author
Debenedictis,
(1997)

Variable
Employment,
Prices and real
GDP for British
Columbia economy
Real Canadian
GDP.

Country
British
Columbia

1992:11995:4

Methodology
Construction and
estimation of
VARmodel

Forecasting
comparisons

49

'.

Findings
VAR yields the
least biased
forecasts for
employment and
British Columbia
GDP.
ARIMA yields the
least biased
estimates for
Canada GDP and
prices.
The result indicates
that 98% and 89%
of the actual data
fell within the 95%
confidence intervals
for ARIMA and
VAR models
respectively.
VAR is a promising
forecasting tool in
the case of British
Columbia.

","'

~'~":""~'"'''''-''''-~~U;J1';'V!~

&!:2:W,fulimi

II

1M

*I

,.

..

&

..

Table 2.1: Summary of Literature Review (Continued)

Data
Author

Variable

Country

Lu, (2009)

GDP

China

Sample Period

Type of Data

Quarterly data

1962-2008

Methodology

50

. -.: c

. "':': :: :

Build an
ARIMAmodel
Chow test to
test for the
presence of a
structural break.

Findings
Data break point
is found
between the 4th
quarter of 1977
and the 1st
quarter of 1978.
The observed
value of the
forecasting
made falls
within the 95%
confidence
interval.

.~

iIIHt*

,,,.._._._.~"""""",","""""'"""~~~.~.,_'~ ._,",,,

ow

iT

t'

,..

Table 2.1: Summary of Literature Review (Continued)

Data
Author
Mittnik &
Zadrozny,
(2004)

Variable
Real GOP

Real industrial
production
Real business
conditions
Expected real
business
conditions

Coun~_ ~aI!!QJe

German

Iyp~()fOat~~.

Period
1970: 1~
2003:4

Quarterly data

1970:1~

Monthly data

2003:12

51

.,
I

rtF

ethodology
Estimated using
VAR models
for in sample
data from
1970: 1 to
1993: 12

Out sample
model data
from 1994: 1
2003:12 is used
to evaluate
GOP forecast.

Findings
Monthly GOP
forecast are
feasible and
produce better
short term GOP
forecast.
Quarterly produce
better GOP
forecast and
aggregation affect
the forecast.
Ifo variables
improve long term
GOP forecast
To estimate larger
models we must
extent to mixed~
frequency
forecasting.
Choosing 0
options and unit
option in
analogous test
improve accuracy
of GOP forecast.

~,~ ~"",,",~v)_~a~

.........

"';'~:'~'~~"':""~~-""."-'-'--

"',,;"

"

II

Table 2.1: SummaI)' ofr.iterature Review (Contiltu~d)

Author
Hoehn, Gruben &
Fomby,
(1984)

Variable
Industrial Production
index
Consumer price
index
Employment
according to survey
business
establishments
Total non
agricultural civilian
employment,
Labour force
Deflated personal
income
Retail sales

Data
Country _ _ Sam~peri()<lTYQf! of Data
Quarterly data
Texas
1969-1980

.1981:11983:2

52

:"':

:-",_'"7:::::

-:-:

Quarterly data

Metho<lology_
ARlMA, Closed
region and
Trickle-down
models.
Out-of-sample
performance of
ARIMA, Closedregion and
Trickle-down
models.
Combination
forecast is
constructed to
compare with
AR1MA.

__Findings
Mean errors for
all variables
except labour
force generate by
two alternative
ARlMA models
were negative.
RMSE also
generally
exceeded the
standard error of
the within sample
often quite
substantially.
Among the seven
series studies, the
two employment
series seem the
most important
for the regional
forecaster to
watch.

he

'?'

e ',"'

'n

,,'

"'Ii"

' i ' "

,.

't

gO 1t1! t!

..

'f'

'Nt

'P;

Table 2.1: Summary ofIAterature Review (Continued)

Author
Schneider &
Spitzer,
(2004)

Variable
National Account
Data
WIFO quarterly
survey
Monthly survey data,
Prices
Foreign trade
Labour market
Financial Variables
Miscellaneous

Data
Country
Austrian

Sample Pe!io~
1988:1
2003:2

~..

IYIle of Data .. _
Quarterly Data
Monthly Data

53

~"

--

..

~Meth()dology

First create a
model by
obtaining the
varying size of an
ordered data set.
The forecasting
performance is
obtained by
performing out
of-sample
forecasts for 30
rolling windows.
Where the
First contains
data from
1991Ql until
2003Q2. The last
three
observations were
omitted for
estimation and
for evaluating
out-of-sample
forecast.

Findings
Broad movements
of the business
cycle are predicted
correctly.
Results from
Pesaran
Timmermann test
show that the
change in the GDP
growth rates
direction is
correctly predicted.
On the other hand,
the forecasting
error for two steps
ahead forecast is
smaller than one
step forecast.
In conclusion,
Factor model
perform better in
small data set, the
number of dynamic
factor could
impacts heavily on
forecast
performance

Ii

...

_.._.- -.,---.----

-,.....

._,..

"

. ._.W"

" . '"

". r

Table 2.1: Summary of Literature Review (Continued)

Data
Author
Krajewski, (2009)

Variable
Polish GDP
Output & Sales
Construction,
Domestic and
Foreign Trade
Prices
Labour Market
Budgetary and
monetary Policy

Country

Sample Period

Type of Data

Poland

1997:1
2008:3

Quarterly Data

Methodology
Dynamic factor
model (DFM) is
used to construct
the economic
indicators and
forecasting.
Forecast accuracy
is compare with
the forecast
accuracy of AR
model and
symptomatic
model.

Finding
From the model,
the RMSE for
DFM is lowest
with only 0.0160
amongARand
causal model
with 0.8116 and
0.0440
respectively.
Changes in data
set influenced
the fmal result
of model
estimation.
In this study it
brought out
increasing
number of
factors and
improvement
estimation
performance.
However, it did
not improve
forecasting
performance.

54

,)"."~""..~, , , , , , , ,., , , -, ", . .....,,-...,,,,,,,",----~~~,

vlte"

"6mb

'''r

{foUN

tn,'ttttttt"

, .

tx1'"

Table 2.1: Summary of Literat!lre Review (Continued)

Author
Kogid, Mulok, Lim
& Mansur,
(2010)

Variable
Log Gross Domestic
Product (LGDP)
Consumption
expenditure (LCE)
Government
expenditure (LG E)
Exports (LX)
Exchange Rate
(LER)
Foreign Direct
Investment (LFDI)

Data
Country
Malaysia

.. Sample Period
1970-2007

Type of Data
Yearly Data

Mtlthodology
Unit root
Co integration
Analysis
Causality
Approach
Error
Correction
Model (ECM)

55

Finding
Consumption
expenditure,
government
expenditure,
export, exchange
rate, and foreign
direct investment
cause economic
growth in the
short run.
Individual test
conducted point
out that only
consumption
expenditure and
export cause
economic growth.
Consumption
expenditure and
export play
important role as
determinant
factors to
economic growth.

'M'

j~

'm

tt

Itt

'I

'j'

' t ' t t n t t n u ; tm

'?tsZWi?':t:'I!:W'rtrtr:nm q

r w

I 'Itt

flttrPlmr

em rt

71'

l ! r lI:n " 1?Ip r.lllllllf1ll1nllllRnlllRlFlbli m T rcF'll

an '

Table 2.1: Summary of Literatllre Review (Continued)

Author
Anaman (2004)

Variable
Annual growth GDP
Annual growth rate
of real total exports
Annual growth rate
of total labour
Annual growth rate
oftotal capital stock
Government size as
the ratio of
government
expenditure to GDP
Lagged ratio of total
investment to GDP

Data
Country
Brunei
Darussalam

Sample Period
1974-2002

Type of Data
Yearly Data

Methodology
Unit root test
Co integration
analysis (ARDL
model)
Unrestricted
Error
Correction
Model

56

-.

Finding
The growth of
export
significantly
influenced longrun economic
growth rates and
the other main
factor is the
relative size of
government where
it influence the
long run growth
rate in the form of
cubic function.
Large government
sizes impeded
economic growth
while moderate
government sizes
enhanced
economic growth

Chapter Three

Methodology
3.0

Introduction

There are a lot of different methods have been used to forecast GDP by other
researchers in the literature review. Although multiple and varying methods have
been used by different authors but the results and objective are the same which is to
forecast the GDP of their own country of study. So, there is no single method that
could satisfy the forecasting needs as every method has its own flows. This paper
would like to propose a time series analysis to identifY the determinant of GDP and
construct a model based on Vector Autoregressive (VAR) model to forecast the GDP
I

of Malaysia.

I,

I:"

This paper begins the empirical testing procedure with model specification in
order to determine which is appropriate to become the determinant of GDP. The
model specification starts from the basic unit root test to examine the stationarity of
the data over a selected period. Then, the testing procedure will extend to Johansen
and Juselius co integration test to determine whether the non-stationary variables
have valid co integration relationships among them. If the variables are co integrated
then it will proceed with Vector Error Correction Model (VECM) modelling.
However, if the variables are not co integrated then Vector Autoregressive (VAR)
modelling will be conducted. The forecast accuracy of the model is checked using
Root Mean Square Error (RMSE), Mean Average Percentage Error (MAPE) and
encompassing test method.
57

"

Chapter 3 consists of the methodology that will be use in this study. This
chapter is organized as section 3.0 which discuss on the introduction. Section 3.1
provides the description of the data used in this study. Then, section 3.2 provides the
empirical testing procedure for this study.

3.1

Model and Data Description

This study adopts the model used by Kogid, Mulok, & Lim (20 I 0) mainly
because the study was conducted in Malaysia and thus making it significant to this
study. The model begins with Yt is a function of Xit which can be specified as below:

(3.1)

Yt = [eXit) ,

where Yt is the logarithm of Gross Domestic Product (LGDP) at time t,

Xit

are log of

broad money (LM2), logarithm of exchange rate (LER), logarithm of fixed deposit
(LFD), logarithm of exports (LX), logarithm of industrial production (LIP) and
logarithm of real household consumption expenditure (LRHCE) at time t, i I , 2,
3, ... , n. Thus, to allow for the inexact relationship between economic variables, the
deterministic economic growth function is modified as follows:

(3.2)

where

Et

is known as disturbance or error. The disturbance term

may well

represent all those factors that affect economic growth but not taken account
explicitly. However, this model ignores Foreign Direct Investment (FDI) as a result
of unavailability of data for Malaysia. Besides that, some relevant variables to GDP

58

I',

that was not emphasize by Kogid, Mulok, & Lim (2010) was also included in the
model.

The variables involve in this study are Gross Domestic Product (GOP), money
supply which is represented by broad money (M2), exchange rate (ER), fixed deposit
rate (FD), exports (X), industrial production (IP) and household consumption
expenditure (HCE) for Malaysia. All the variables estimated will be in logarithm
form. This is mainly because according to the model by Kogid et a1. (2010), all the
variables used are in logarithmic form. Moreover, the variables will be in logarithm
form because the differences in some of the variables values are large and there are
also some values that are in percentage form thus using logarithmic form could
,,.

lower the differences in values.

Broad money (M2) is positively correlated with GDP which is illustrated by

Arnold (2008). He state that a change in the aggregate demand and thereby change
the price level and also GDP in the short run. This is to say that an increase in the
money supply would shift the demand curve to the right which will move the
economy to a higher point. On the other hand, if money supply is decrease then
lower level of GDP will be produce.

The changes in exchange rate (ER) would eventually affect the stock prices.
Stock prices on the other hand could influence the changes in GOP as stock price
reflect investment (Franck and Young, 1972; Bartov and Bodnar, 1994; Fernandez,
2006). The lower the stock price indicates the lower the investment as investment is
lower than the output is also affected as investment contribute to the country GOP.
Therefore, exchange rate is negatively correlated with GDP.
59

Fixed deposit rate (FD) represents interest rate. The relationship between
interest rate and GDP was portrayed by Obamuyi (2009) that imply the behaviour of
interest rate is important in economy growth that normally represented by GDP or
GDP growth as interest rate could affect investment and investment could affect the
output of the country. In other words lower interest rates reduce the value of a
country's currency which in term lowers the prices of goods and services sold for
export for that particular country which leads to greater spending for that country's
goods and services thus interest rate and GDP is negatively correlated.

When the demand of exports (X) for other foreign countries increases thus
directly it could increase the income of the country whereas indirectly is through the
changes of the term of trade. Exports also would likely to alleviate foreign exchange
constraint thus provide greater access to international market. Some studies that
show the relationship between export and real GDP are Dritsaki, Dritsaki &
Adamopoulos (2004), Sentsho (2000) and Awokuse (2002). Therefore, export and
GDP is positively related to each other.

Indirectly the changes in real activity or Industrial production (IP) could affect
the interest rate that could cause a decrease or increase in future production and
which according to Peiro (1996), interest rate could affect the stock price through
variation of future product which could influence the country output. Therefore, real
activity or industrial production is positively related to GDP.

According to Kogid, Mulok & Lim (2010), consumption expenditure in this

sense household consumption expenditure (HCE) is one of the variables that play an
important role as a determinant factor to a country economic growth in this sense
60

GOP in Malaysia. There are also a number of studies done regarding the relationship
between consumption expenditure and economy growth such as FoIster and
Henrekson (1999) and Kweka and Morrissey (1998). Moreover, according to Keynes
(1936), consumption expenditure is positively correlated with GOP.

All the data used in this paper are retrieved from various issues of International
Financial Statistic (IFS) published by International Monetary fund (IMF). This study
uses yearly data and the periods of data involve in this study covers from 1970 to
2010. However, the data is divided into two sub-samples where the first sub-sample
is from 1970 to 2008 also known as in-sample. In-sample is used to forecast the
GOP. On the other hand, the second sub-sample is from 2009 to 2010 also known as
out-of-sample. Out-of-sample is kept for evaluation of the forecast accuracy.
Because data is unavailable for the year earlier than 1970 for certain variables so the
data starts from 1970. In addition, year 2009 is used as the start of the out sample
because according to the trend of GOP, Malaysia is affected more in 2009 than the
financial crisis in 1997 so this study would like to study whether the GOP for the
earlier year can predict the GOP for the crisis in 2009.

3.2

Empirical Testing Procedure

This paper starts by identifying the determinants of GOP and run diagnostic
checking on the model to make sure that it is well fit and free from econometric
problems and also in order to determine the order of integration for the variables as
only when LGDP and other variables are integrated at 1(1) then can proceed to
undergo Johansen Multivariate cointegration test.

61

3.2.1 Unit Root Tests

For this study, two types of unit root test would be conducted which are
Augmented Dickey-Fuller (ADF) test and Kwiatkowski-Phillips-Schmidt-Shin
(KPSS) test. In the ADF test, the variables of study in level are first estimated in the
form of either with trend intercept or with intercept but without trend. Then in first
difference, the variables are in the form of either intercept or none. The reason that it
is done as stated in the sentence previously because estimation for level form
actually shows the long run estimation of the variables while the first differences
captures the short run estimation of the variables as it does not include trend. The
variables are also estimated the same way for PP test.

The first thing to do in order to run unit root test is to estimate all the variables
in the first level for the form with trend and intercept to identifY the significance of
the trend for that particular variable. The result is then reported in a table for the
variables that are significant. The estimation will continue to be conducted until
arrive to the stage that all the variables are stationary. ADF test and PP test null
hypothesis state that the variables contain unit root or in other words the variables
are not stationary while the alternate hypothesis for these two tests state that the
variables are stationary. (Rapach and Weber. 2004)

3.2.1.1

Augmented Dickey-Fuller (ADF)

Augmented Dickey-Fuller (ADF) test is developed by Dickey and Fuller

(1979). This test is conducted by augmenting the three equations (3.3), (3.4) and

62

(3.5) by adding lagged values of the dependent variables ~Yt. After augmenting the
three equations, equation (3.6) is obtained. The equations involve are as below:

(3.3)

(3.4)

(3.5)

(3.6)

where Ct is a pure white nOIse error tenn, ~Yt-l

3.2.1.2

= (Yt - 1 -

Yt - 2), ~Yt-2

Phillips-Perron (PP)

Phillips-Perron unit root tests were developed by Phillips and Perron (1988) in
which it differs from ADF mainly in the way they deal with serial correlation and
heteroskedasticity in the errors. PP tests ignore any serial correlation in the test
regression. The test regression for PP test is

~Yt

= pi Dt + rcYt-l + Ilt ,

(3.7)

where Ilt is 1(0) and may be heteroskedastic. Any serial correlation and
heteroskedasticity in error tenn can be corrected by modifying the test statistic
trr=o and Tft. The modified test statistics which are denoted as Zt and Zrr are:

Zt

2
(8l2 )1/2 . tn=o _~ (l2-8
) (T'SE(ft))
l2
82
2

63

(3.8)

_ T"
1 r2 'SE(rt) ( A1'2
ZtTr-2
82

"2)

(3.9)

-(}"

3.2.2 Cointegration Test

Johansen and Juselius (1990) multivariate cointegration test is used to

investigate the long run relationship between GDP and its determinant. The lag
structure is determined automatically by the statistical package. The core movement
in the long run can be defined as follow:

(3.10)

Where,

Mi

= -(I -

ill - ... - ilD ,(i = 1, ... , k - 1)

And

Model (1) is expressed as a traditional first difference of a VAR-model except

for /IX t - k , thus coefficient matrix il is used to be investigated to find out whether it
contains information about long-run relationship among the variables in the data.
The rank of il depends on two likelihood ratios which include Trace test and
Maximum Eigenvalue test. Trace test as expressed as:

Ttrace

=- T Lf=r+l1n (1 -

~) ,

(3.11 )

64

Where, T denotes the number of valid observations for estimation use and ~ is the
jth largest

estimated eigenvalue.

On the other hand. Maximum Eigenvalue is expressed as:

Amax = -T In (1- ~+1)

(3.12)

Where, T denotes the number of valid observations for estimation use and ~+1 is the
largest eigenvalue at r

+ 1.

After the variables in the equation are prove to be stationary and cointegrated
with each other then the next step is to forecast the model. The next section shows
modelling technique used in this study to forecast the model.

3.2.3 Forecasting Models

3.2.3.1

ARIMAModel

One of the forecasting models in this study is obtain using ARIMA model.
ARlMA is also known as Auto Regressive Integrated Moving Average intends to
describe the current behaviour of the variables with their past values. This model was
introduce by Box and Jenkins (1984) and can also be call as Box-Jenkins model.
ARlMA has two parts in which Integrated (I) represent the amount of differencing to
be performed on the series to be stationary and the second component is ARMA
which could be decomposed into AR and MA components. Autoregressive (AR)
captures the correlation between current value of the time series and some of its past

65

values while Moving Average (MA) represents the duration of the influence of a
random shock. ARlMA model can be written as:

(3.13)

where Yt represents the variables that to be forecasted in the model in this case it is
gross domestic product and at is the random noise.

3.2.3.2

Fundamental Models

The fundamental models obtain for forecasting the GDP is by estimating the
dependent variable and independent variables using Ordinary Least Square Method
(OLS). The fundamental models is made up from three models in which one consists
of the full model, the second consist of only the variables that are significant in the
first model and the third model is obtain from by regressing GDP with the variable
with the highest correlation among all the other variables.

3.2.3.3

Random Walk Model

The next model that will be use to forecast the GDP is random walk model or
also known as naive model which had been used by a lot of people to forecast certain
variables (Agwuegbo, Adewole & Maduegbuna, 2010; Pesaran & Oick, 2008). The
model for the purpose of this study is computed as:

(3.14)

where t is the year estimated and t - 1 is a year before the estimated year. Y would
be the variable estimated in the study in which in here its LGDP.

66

3.2.4 Model Specification and Diagnostic Checking

Nonnally models are check using diagnostic checking in order to see whether
autocorrelation and heteroscedasticity existed however when estimating the models
Newey-West estimator is used where the standard assumption of regression analysis
do not applies thus the problem of autocorrelation, heteroscedasticity in the error

tenn could be overcome. The variables is then checks for nonnality to detennine
whether the variables is nonnal in order to select which type of correlation

f
I

coefficient to be used in testing for correlation.

1!
,

3.2.4.1

Normality Test

Nonnality test is to check whether the error tenn is nonnally distributed. The
test that is adopted to check for the nonnality of the error tenn is Jarque-Bera (JB)
Test. According to Jarque and Bera (1987), JB test is an asymptotic or large sample
test. This test computes the skewness and kurtosis measure of the OLS residuals
using the following test statistic:

JB

= n [52 + (K-3)2]
6

(3.15)

24'

Where n is the sample size, S is the skewness coefficient, K is the kurtosis

coefficient and for nonnally distributed S

3.2.5.2

= 0 and K = 3.

Correlation Test

According to Bolboaca and Jantschi (2006), correlation coefficient is a simple

statistical measure of relationship between the dependent and one or more

67

independent variables. This test is used in finding out the impact of each variable
toward GDP and also to investigate whether the expected sign is according to the
theory. Moreover, correlation test is important in determining the variables that had
the highest correlation to be regress with GDP. There are three types of correlation
coefficient would be considered depending on the result of the normality test.

The first correlation coefficient is Pearson correlation coefficient which

requires the variables to be normally distributed. Pearson coefficient can take values
from -1 to + 1 in which +1 show that the variables are perfectly linear related by an

I
I
~

increasing relationship and a value of -1 show that the variables are perfectly related
by a decreasing relationship. In addition, value of 0 shows that the variable are not
linearly related to each other. The test statistic for Pearson is given by:

Ho:

fPrs

(3.16)

(3.17)

where equation 3.16 indicates that there is no correlation between the variables and
equation 3.17 indicates that the variables are correlated.

The second correlation coefficient would be considered is Spearman's rank

correlation coefficient. This is a non-parametric measure of correlation between
variable that will explore deeply to how well arbitrary monotonic function could
describe the relationship between the two variables. Spearman's in overall is
satisfactory for testing the null hypothesis when no relationship exist however it is
difficult to interpret as a measure of the strength of relationship. The values for
Spearman's rank correlation is similar to Pearson but for Spearman's there are two
68

type of correlation present in which the first one is Spearman's rank correlation
coefficient which is given by:

Ho:

fS pm =

HI:

fS pm <>

(3.18)
0

(3.19)

On the other hand, the second correlation coefficient for Spearman's

IS

semI

quantitative correlation coefficient which is given by:

Ho: fsQ= 0
HI:

fsQ<

(3.20)

>0

(3.21)

Where equation 3.18 and 3.20 indicates that there is no correlation exists between
the ranked pairs while equation 3.19 and 3.21 indicates that the ranked pairs are
correlated. The third correlation coefficient considered is Kendall's rank correlation
coefficients. Kendall-tau is one of the non-parametric correlation coefficient that is
used to test correlation between non-interval scaled ordinal variables. Kendall's tau
is supposedly equivalent to Spearman rank correlation coefficient. Kendall's tau
consists of three correlation coefficients that are known as tau-a, tau-b and tau-c
respectively. The statistical significance of Kendall's tau is given as:

Ho:

iKen,a=

HI:

iKen,a<>

(3.22)

(3.23)

Where equation 3.26 means that there is no correlation between the two variables
and equation 3.27 indicates that the two variables are correlated. The same statistical

69

significance also applies to the other two type of Kendall's tau correlation
coefficient

3.2.6 Forecast Evaluation Criteria

After forecasted the model, the accuracy of the model is evaluated using the
last 2 years which is from 2009 to 2010 using the common forecast evaluation
criteria.

3.2.6.1

Root Mean Square Error (RMSE)

RMSE is the most frequently used measure and is known to be more sensitive
to outliers. The general form ofRMSE is found in equation (3.18).

RMSE =

(n~l) [Lf=l(LlY! -

(3.24)

LlYt)]2 ,

where n is the number of quarters between the periods, LlY! is forecasted LlY and LlYt
is the actual LlY at period t.

3.2.6.2

Mean Absolute Percentage Error (MAPE)

MAPE is another test that will be use to evaluate the accuracy of the forecast.
The test statistic is written as:

2. Labs (At+n-tFt+n))
(N
At+n

100 ,

(3.25)

where abs stands for the absolute value. Note that, At+n denotes the actual value of
a specific variable in period t

+ n and t Ft+n is the forecast made in period t for t+n.

70

Chapter Four

Results and Discussion

4.0

Introduction

The aim of this study is to find the determinants of GDP in Malaysia and later
the determinants identified would be used to forecast the GDP of Malaysia. From the
previous study, it is known that there are lots of models used to predict the GDP of a
country. These models consist of fundamental models (Vector Autoregressive
(VAR) model, Bayesian Autoregressive (BVAR) model, Dynamic Factor model,
Dynamic Stochastic General Equilibrium (DSGE), etc.) and Autoregressive
Integrated Moving Average (ARIMA) time series models.

In this study, few models had been estimated and they are ARIMA models,
fundamental models and random walk modeL ARIMA model is used in order to
describe the current behaviour of the GDP with its own past values. This means that,
ARIMA is estimated to determine whether past values of gross domestic product
(GDP) could be used to forecast the future value of GDP. Besides, fundamental
models are used in order to determine whether the inclusion of other external factor
besides GDP could increase the accuracy of the forecast. In other words, in order to
investigate if other factors could provide useful insights on the future movement of
the GDP of Malaysia. Lastly, random walk model is used to examine whether last
period's GDP could forecast this period GDP or this period GDP could forecast next
period's GDP with satisfactory accuracy. Besides, random walk was used as a
comparison with ARIMA and fundamental models because if random walk models
71

forecast accuracy is better than the other two models then government would not
have to waste their time modelling a model to forecast GDP as they could always
forecast it from the previous year.

The full sample annual data used for this study range from 1970 to 2010 in
which the in sample is from 1970 to 2008 and the out sample is from 2009 to 2010.
The time series models used in this study is represented by ARIMA models. In order
to determine whether ARMA or ARIMA model needs to be used, it is a must to
identify the order of integration, which could be tested using unit root tests.
Meanwhile, the first step in fundamental modelling is to make sure that all the
variables are integrated at the same order. Similar to ARIMA, to test the order of
integration, unit root tests are carried out. The next step is to make sure that all the
variables used are cointegrated and this can be investigated using Johansen
cointegration test. However, before Johansen cointegration test can be conducted, the
variables must be integrated at order one, 1(1). Only after the variables are identified
to be cointegrated then we have the basis to estimate the fundamental models.

On the other hand, random walk model does not need to first run unit root tests
and cointegration test as random walk model just involve comparing the previous
year data with the current year data. After that the forecast accuracy of all the models
will be computed for comparison purpose. All the models' accuracy will be
determined with Root Mean Square Error (RMSE) and Mean Absolute Percentage
Error (MAPE). This criteria are explained in chapter 3.

72

4.1

Unit Root Test

As stated in the previous chapter, Augmented Dickey Fuller (ADF) and

Phillips-Perron (PP) unit root tests are applied to determine the order of integration
for variables in forecasting Malaysia GDP and also the stationarity of the data. The
optimum lag length for unit root test is chosen based on Schwartz Information
Criteria (SIC). SIC is similar to Akaike Infomation Criteria (AIC) only that AIC is
aimed at finding the best approximating model to the unknown data generating
process while SIC is designed to identify the true model (Cavanaugh and Neath,
1999). SIC is derived within a Bayesian framework. It reflects sample sizes and has
properties of asymptotic consistency. Moreover, according to Tweneboah and Adam
(2008) SIC is also preferred as it typically chooses shorter lags compared to AIC. In
this study, unit root testing will be conducted for the full sample which is from 1970
to 2010 and also for the in sample which is from 1970 to 2008.

4.1.1 Unit Root Test Result

Table 4.1 shows the results of the ADF tests for the whole sample, which range
from 1970 to 2010. For ADF test, null hypothesis will be rejected if the test statistic
value is more than the critical value in magnitude. Referring to Table 4.1 for the
column for ADF test result, it can be seen that, we cannot reject the null hypothesis
for the variables logarithm of gross domestic product (LGDP), logarithm of broad
money (LM2), logarithm of fixed deposits (LFD), logarithm of exchange rate (LER),
logarithm of export (LX) and logarithm of industrial production (LIP) basically
because the magnitude of the test statistics have a lower value compared to the
magnitude of the critical values for 10 percent, 5 percent and 1 percent. However,
73

J.

1,
the null hypothesis can be rejected in the first difference for LODP, LM2, LFD,
LER, LX and LIP at 10 percent or better significance level as a result of the higher
value of the magnitude of the test statistics compared to the magnitude of the critical
values with the exception of LIP. For LIP test statistic is smaller than its critical
value magnitude for 10 percent, 5 percent and 1 percent thus the null hypothesis is
not rejected. However, the null hypothesis can be rejected at 1 percent significant
level for this variable in the second difference form. As a whole, the ADF tests for
the full sample indicates that LODP, LM2, LFD, LER, LHCE and LSP are integrated
of order one, I (1). Meanwhile, for LIP it is integrated at order two, I (2).

On the other hand, for PP test results which are shown in Table 4.2 suggest
that LODP, LM2, LFD, LER, LX, LIP and LHCE are not stationary at their level
form. This is because we could not reject the null hypothesis even at 10 percent as
the magnitude value of the test statistic for these variables are lower compared to the
magnitude values of their 10 percent critical value.

As for the first difference for all the variables involved (LODP, LM2, LFD,
LER, LX, LIP and LHCE) we could reject the null hypothesis for 10 percent or
better. Overall, PP test results for the full sample indicate that LODP, LM2, LFD,
LER, LX, LIP and LRHCE are integrated at order one, 1(1).

Table 4.3 shows the results of the unit root test for ADF for in sample which is
from 1970 to 2008. Following the same principle of interpretation as before, the
ADF tests results for the in sample indicates that LODP, LM2, LFD, LER, LRX and
LHCE are integrated of order one, I (1). Meanwhile LIP is integrated at order two,
1(2).
74

~-,

I
",
j

Table 4.4 shows the results of the unit root test for PP for in sample which is
from 1970 to 2008. Similar to the interpretation technique used for the full sample
PP test, the results of the PP test for in sample indicate that LGDP, LFD, LM2, LER,
LIP, LX and LHCE are integrated at order one, 1 (1).

From ADF test results for full sample and in sample, both of these tests yields
consistent results in which both also state that LGDP, LM2, LFD, LER, LX and
LHCE is integrated at order one, 1(1) while ADF test for both sample also identifY
LIP as integrated at order two, 1(2). On the other hand, PP test results show that for
both the full sample and in sample unit root testing, LGDP, LM2, LFD, LER, LX,
LIP and LHCE is integrated at order one, 1(1). As the estimation will be conducted
for in sample it is important to ensure that all the variables during this period are
stationary. Judging from Table 4.3 and Table 4.4 which constitute the period of in
sample, ADF and PP unit root tests indicate similar results of level of integration for
all the variables with the exception of LIP. The final decision on whether LIP is I( 1)
or 1(2) is based on PP test. Phillips and Perron (1988) also state that PP test has
significant advantages when there are moving average components in the time series
and could offer a promising alternative to ADF. Moreover, according to to Choi
(1992), PP tests show better finite sample performance than ADF test. Furthermore,
PP test is also more robust to general forms of heteroskedasticity in the error terms
(Onour, 2007). Based on PP test that is more robust than ADF test, this study
concludes that for in sample, LGDP, LFD, LM2, LER, LX, LIP and LHCE are all
integrated at order one, I( 1). The same conclusion can be made for the full sample
where LGDP, LFD, LM2, LER, LX, LIP and LHCE are all integrated at order one,
1(1).
75

--~
Table 4.1: ADF results for 1970 to 2010
ADF

Variables
None

LGDP
LM2
LFD
LER
LX
LIP
LHCE

First Difference

Level
Intercept

Intercept & Trend

None

-2.109(0)
-2.434(1)

Second Difference
Intercept

None

-5.807(0)***
-2.353(1)**

-3.046(1)
-2.703(0)
-1.378(0)
-1.685(0)

-5.215(0)***
-6.201(0)***
-1.634(3)*
-1.183(3)

-2.991(1)

-7.332(2)***
A.859(0)***

Finding

1(1)
1(1)
1(1)
1(1)
1(1)
1(2)
1(1)

Notes: LGDP= logarithm of Gross Domestic Product, LM2= logarithm of Broad money, LFD= logarithm for Fixed Deposits, LER= logarithm of exchange rate, LX= logarithm log
of exports, LlP= logarithm of industrial production, LHCE= logarithm of household consumption expenditure. Asterisks (*"'*), (*"') and (*) denote statistical significant at 1%, 5%
and 10% level. Figures in the bracket 0 are the optimal lag length. The period of estimation is from 1970 to 2010.

76

----.. -----------~

!Il'~"-

Table 4.2: PP results for 1970 to 2010

PP

Level

Variables
None

LGDP
LM2
LFD
LER
LX
LIP
LHCE

Intercept

First Difference
Intercept & Trend

None

*2.115(1)
*2.298(1)
-2.240(5)
-2.713(1)

Finding
Intercept

*5.796(2)*"'*
-6.086(2)*"'*
-5.701(12)***
-6.204(4)***

-1.677(6)
-1.714(1)

-4.089(4)***
-2.689(3)***

1(1)
l(l)
1(1)
1(1)
1(1)
1(1)

-2.401(3)
-4.773(3}*~*
Notes: LGDP= logarithm of Gross Domestic Product, LM2= logarithm of Broad money, LFD= logarithm for Fixed Deposits, LER= logarithm of exchange rate, LX= logarithm log
of exports, LIP= logarithm of industrial production, LHCE= logarithm of household consumption expenditure. Asterisks (U.), (U) and (oil) denote statistical significant at J%, 5%
and 10% level. Figures in the bracket 0 are the optimal lag length. The period of estimation is from 1970 to 2010.

77

jib.

Table 4.3: ADF results for 1970 to 2008

ADF

Variables
None
LGDP
LM2
LFD
LER
LX
LIP
LHeE

First Difference

Level
Intercept

Intercept & Trend

None

-2.828(1)
-2.135(1)

Second Difference
Intercept

None

-5.274(0)***
-2.222(1)**

-3.122(1)
-2.905(0)
-2.590( 1)
-1.174(0)

-4.876(0)***
-6.162(0)***
-6.334(0)** *
-1.105(3)

-2.982(U

-7.198(2)***
-4.6Q5(Q)***

Finding

1(1)
I(l)
1(1)
1(1)
I( 1)
1(2)
1(1)

Notes: LGDP= logarithm of Gross Domestic Product, LM2= logarithm of Broad money, LFD= logarithm for Fixed Deposits, LER= logarithm of exchange rate, LX= logarithm log
of exports, LIP= logarithm of industrial production, LHCE= logarithm of household consumption expenditure. Asterisks (***), (**) and (*) denote statistical significant at 1%, 5%
and 10% level. Figures in the bracket 0 are the optimal lag length. The period of estimation is from 1970 to 2008.

78

~,"j

~:.-

Table 4.4: PP results for results for 1970 to 2008

PP

Variables

None
LGDP
LM2
LFD
LER
LX
LIP
LHCE

First Difference

Level

Intercept

Intercept & Trend

None

-2.295(0)
-2.432(1)

Finding
Intercept
-5.245(2)***
-6.064(2)***

-1.937(3)

-5.310(11)***
-2.885(3)
-2.434(1)

-6.194(5)***
-6.447(2)***

-1.353(5)

-2.470(3)**

1(1)
1(1)
1(1)
1(1)
l(l)
l(l)
1(1)

-2.427(:;n ~. _ _
-4.515(3)***
Notes: LGDP= logarithm of Gross Domestic Product, LM2= logarithm of Broad money, LFD= logarithm for Fixed Deposits, LER= logarithm of exchange rate, LX= logarithm log
of exports, LIP= logarithm of industrial production, LHCE= logarithm of household consumption expenditure. Asterisks ("'U), (*"') and ("') denote statistical significant at 1%, 5%
and 10% level. Figures in the bracket 0 are the optimal lag length. The period of estimation is from 1970 to 2008.

79

'1

4.2

ARIMA Model

From the unit root test result conducted, according to PP test result, LGDP is
stationary at 1(1) so ARIMA models instead of ARMA models should be estimated
with the amount of four lags chosen for the autoregressive (AR) and moving average
(MA) component. The models' in-samples forecast accuracy and out-samples
forecast accuracy will then be evaluated to indicate the best ARIMA models. Two
types of ARlMA models will be estimated in which both of them also consist a total
of 4 lags the only difference between these two types are that one is with constant
and another is without constant. There are all together 24 models each for ARlMA
models with constant and ARlMA models without constant. This is done in order to
compare which types of ARIMA models could give better forecast accuracy.

4.2.1 ARIMA Model Forecasting Accuracy

Table 4.5 shows the forecasting performance of ARlMA model with constant.
From Table 4.5, it is found that the ARIMA model with the lowest RMSE for in
sample is ARlMA (4,1,3) model in which this model also produces the lowest
RMSE value for out of sample forecast. Meanwhile, the model that yields the lowest
MAPE result is also ARIMA (4,1,3) in which this particular model also produces the
lowest MAPE result for out of sample forecast.

Table 4.6 shows the forecasting performance of ARIMA without constant. The
smallest MAPE value for the in-sample period is ARlMA (2,1,3) but the smallest
out-of sample value for MAPE falls to ARlMA (0,1,4). On the other hand, the
smallest value ofRMSE for in sample falls to ARlMA (2,1,3). So, the best in sample
80

model for ARIMA model without constant is ARIMA (2,1,3) although it do not
actually produces the best out of sample forecast.

From the results of the best models from both of the tables, it can be seen that
ARIMA (2,1,3) without constant did not yield better result than ARIMA (4,1,3) with
constant. Moreover, the RMSE and MAPE for ARIMA (4,1,3) yield more consistent
results compare to ARIMA (2,1,3) without constant. Thus, the model that could
perform the best for full sample, in sample and out sample is ARIMA (4,1,3) model
where in full sample, this model had the lowest RMSE and MAPE value with 0.0854
and 0.6007 respectively compared to the other 24 models. On the other hand, for in
sample, ARIMA (4,1,3) model also had the lowest RMSE and MAPE values of
0.0786 and 0.5657 respectively compared to the other ARIMA with constant modeL
In addition, the out-of sample RMSE and MAPE values of 0.1416 and 1.0492 for
ARIMA (4,1,3) model and the lowest values among all ARIMA with constant
models.

81

Table 4.5: Forecasting Perfonnance of ARlMA models with constant (Yearly)

ARlMA
(0,1,1) with constant
(0,1,2) with constant
(0,1,3) with constant
(0,1,4) with constant
(1,1,0) with constant
(1 , 1, 1) with constant
(1,1,2) with constant
(1,1,3) with constant
(1, 1,4) with constant
(2,1,0) with constant
(2,1,1) with constant
(2,1,2) with constant
(2,1,3) with constant
(2,1,4) with constant
(3,1,0) with constant
(3, 1,1) with constant
(3,1,2) with constant
(3,1,3) with constant
(3,1,4) with constant
(4,1,0) with constant
(4,1,1) with constant
(4,1,2) with constant
(4,1,3) with constant
(4,1,4) with constant

__ ~Full
RMSE
0.2019
0.2019
0.1968
0.1573
0.2137
0.3288
0.1989
0.1877
0.3683
0.2229
0.2288
0.1590
0.2297
0.2191
0.1318
0.1139
0.1542
0.0858
0.1048
0.1076
0.1261
0.1385
0.0854
0.0869

~ple

In-sample
RMSE
MAPE
1.4246
0.2023
0.2023
1.4246
0.1956
1.3835
0.1486
1.0326
0.2146
1.5310
0.3363
2.7250
1.4193
0.1978
0.1857
1.3163
0.3756
3.0708
0.2245
1.6230
0.2229
1.6800
0.1559
1.0587
0.2315
1.6292
0.2164
1.5453
0.1275
0.8702
0.1086
0.7488
0.1425
0.9915
0.0795
0.5747
0.6899
0.0974
0.1 0 11
0.7017
0.1222
0.8212
0.1349
0.8985
0.5657
0.0786
0.0795
0.5745

MAPE
1.4253
1.4249
1.3949
1.0825
1.5266
2.6323
1.4292
1.3326
2.9850
1.6118
1.6818
1.0837
1.6197
1.5661
0.9000
0.7816
1.0541
0.6076
0.7296
0.7392
0.8496
0.9262
0.6007
0.6120

Out-sample
RMSE
MAPE
0.1960
1.4529
0.1942
1.4389
0.1954
1.4467
0.1804
1.3358
0.1948
1.4441
0.1914
l.4158
0.2089
1.5441
0.1978
1.4652
0.1566
1.1596
0.1929
1.4295
0.2125
1.5717
0.2202
l.6327
0.2334
1.7300
0.2400
1.7787
0.1844
1.3654
0.1800
1.3325
0.1894
1.4013
1.1636
0.1585
0.1903
1.4093
0.1781
1.3181
0.1922
1.4224
0.1897
1.4054
0.1416
1.0492
0.1887
1.3941

Notes: Full sample cover the range from 1970 to 2010, In-sample from 1970 to 2008 and Out-of sample from
2009 to 2010. The estimation period is from 1991 to 2008 while the forecasting period from 2009 to 2010.

82

Table 4.6: Forecasting Performance of ARIMA models without constant(Yearly)

ARIMA

(0,1,1) without constant

(0, I ,2) without constant
(0,1,3) without constant
(0,1,4) without constant
(1,1,0) without constant
(1,1,1) without constant
(1,1,2) without constant
(1,1,3) without constant
(1,1,4) without constant
(2,1,0) without constant
(2,1,1) without constant
(2,1,2) without constant
(2,1,3) without constant
(2,1,4) without constant
(3,1,0) without constant
(3,1, I) without constant
(3,1,2) without constant
(3,1,3) without constant
(3,1,4) without constant
(4,1,0) without constant
(4,1,1) without constant
(4,1,2) without constant
(4,1,3) without constant
(4,1,4) without constant

Full sam~
RMSE
MAPE
2.6159
19.0622
2.5969
18.8813
2.6593
19.4787
2.1081
14.3927
2.3487
16.8325
0.2339
1.6713
0.2474
1.7504
0.2248
1.6046
0.2466
1.7389
2.0733
14.7677
0.2418
1.7597
0.3115
2.4030
0.2036
1.5456
1.8761
13.2499
0.6655
3.7875
0.2406
1.5877
0.4308
2.9534
0.3118
2.1870
0.3543
2.4330
0.3833
2.2542
0.6835
3.9742
1.0222
6.3865
1.2086
7.6003
0.2229
1.4872

In-sample
RMSE
MAPE
2.5152
18.4729
2.4962
18.2908
2.5585
18.8925
2.0084
13.7855
2.2490
16.2436
0.2259
1.6214
0.2392
1.6978
0.2179
1.5606
1.6861
0.2384
1.9770
14.2031
0.2368
1.7266
0.3011
2.3583
0.2000
1.5256
1.7851
12.7] 54
0.6045
3.4362
0.2249
1.4969
0.4090
2.8210
0.3041
2.1322
0.3378
2.3307
0.3408
2.0307
0.6228
3.6210
0.9463
5.9369
1.1230
7.0871
0.1931
1.3520

Out-sample
RMSE
MAPE
0.1139 0.7213
0.1210 0.8136
0.1805
1.3070
0.0537 0.3928
0.1763
1.3004
0.1484
1.0909
0.1552
1.1411
0.1440
1.0483
0.1515
1.1085
0.1852
1.3711
0.1451
1.0641
0.2]59
1.5996
0.1571
1.1319
0.2104
1.5558
0.1769
1.3087
0.1682
1.2419
1.1321
0.1542
0.1399
1.0147
0.1537
1.1250
0.1692
1.2504
0.1853
1.3709
0.1821
1.3481
0.1332 0.9790
0.2510
1.8608

Notes: Full sample cover the range from 1970 to 2010, In-sample from t 970 to 2008 and Out-of sample from
2009 to 2010. The estimation period is from 1991 to 2008 while the forecasting period from 2009 to 2010.

4.2.2 Comparison between yearly and quarterly ARIMA models

Apart from forecasting the yearly LGDP with constant AR1MA models, this
study also attempts to analyze the quarterly data from 1991 quarter 1 to 2010 quarter
4. Thus, the forecasting accuracy for yearly LGDP data from 1970 to 2010 is
compared to quarterly LGDP data from 1991Ql to 201OQ4. Because of the
unavailability of quarterly data for GDP for years earlier than 1991 thus only 1991
quarter 1 data onwards will be used for GDP quarterly GDP. According to PP test
results (refer to Appendix A, Table Al and A2), quarterly LGDP is integrated at
order zero, 1(0) as the test statistic in the level stage is more than the 10 percent
83

..,

critical value thus the null hypothesis is rejected. Because quarterly LGDP is
integrated at 1(0) thus ARMA model is used and the results can be seen in Table B 1
in Appendix B. From the table, it is found out that the ARMA model that yield the
lowest RMSE result for in-sample forecast is given by ARMA (1,3) model with
constant. However this model did not produce the best forecast for out-of sample
forecast. Nevertheless, ARMA (1,3) did produce the second best out of sample
forecast for RMSE. On the other hand, for MAPE, the lowest value is also given by
ARMA (1,3) model but, different from RMSE criterion. The best out of sample
forecast for RMSE is ARMA (4,4) while for MAPE is ARMA (2,1). Based on all the
ARMA models, the best model selected is based on RMSE criterion in which
ARMA (1,3) model is chosen as the best model.

By companng quarterly ARMA (1,3) model with yearly ARlMA (4,1,3)

model. ARMA (1,3) model tend to produce better full sample, in sample and out of
sample forecast compared to yearly data. Thus, in other words, forecasting GDP with
quarterly data actually yield better result than forecast using early data. This finding
is interesting as it always desirable for one to be able to forecast GDP with higher
frequency.

4.3

Fundamental models

Time series model such as ARMA and AR1MA models might be a good way
to forecast GDP. However, as ARMA and ARIMA models had been built to explain
the past, they will tend to be biased toward the past in the sense that they will weight
past information more heavily than new information and will perform poorly (Wei &
Awazu, 2008). Thus, a fundamental model is needed to include all the other
84

determinants of GDP to forecast GDP. This is because GDP does not only come
from the previous values but also as an outcome of trading, investing and spending
of other external sectors.

There will be three models that are considered in order to select the best
fundamental model that could forecast the GDP with the highest accuracy. The first
model is the inclusion of all the variables that are integrated at order one, I (1)
according to the unit root test results discussed in the previous section. LGDP serves
as the dependent variable, while, LM2, LER, LFD, LX, LIP and LHCE serves as the
independent variables. This model will be labelled as Modell.

The second model only includes variables that are significant from Model 1
and thus this second model will be label as Model 2. Meanwhile the third model is
obtain using bivariate correlation in which only the variable with high correlation
value to LGDP will be estimated and the third model will be name as Model 3. To
estimate both of these models, Newey-West HAC standard errors and Covariance is
used because if this method is used then autocorrelation and heteroscedasticity
would not be an econometric problem in the regression. In other words, the t
statistics are from the presence of autocorrelation and heteroscedasticity. The in
sample periods will be estimated. Before estimating both of the models,
cointegration test will be conducted first to ensure that the model is cointegrated thus
Johansen Multivariate Cointegration test will be used. The results obtained from the
cointegration test and estimation of both the models are presented in the sections
4.3.1,4.3.2 and 4.3.3.

85

4.3.1 Modell

For Johansen cointegration test, in order to reject the null hypothesis, the test
statistic must be larger than the critical value. As shown in Table 4.7, for the null
hypothesis of r = 0, the trace test is 179.241 and it is larger than the 5 percent critical
value of 125.615. Thus, we can reject the null hypothesis. On the other hand, for the
same null hypothesis the test statistic for max eigenvalue is 78.658 and it is also
larger than its 5 percent critical value of 46.231, thereby resulting in rejection of the
null hypothesis. For the null hypothesis of r ::; 1 the test statistic for trace test is
100.583 which are larger than the 5 percent critical value of 95.754. This leads to the
rejection of the null hypothesis. Consistently, for the same null hypothesis the test
statistic for max eigenvalue of 43.113 is also larger than its 5 percent critical value of
46.231 thus rejecting the null hypothesis. From r

:s

2 onwards, the entire test

statistics for trace and max eigenvalue are smaller than their 5 percent critical values
so the null hypothesis cannot be rejected. This means that the cointerating vector is
not more than 2. As a result, it can be concluded that Model 1 has two cointegrating
vectors which means that the variables in Model 1 are co integrated with one another
thus showing that long run relationship exist. Model 1 is then estimated to identifY
the variables that are significant and also for forecasting accuracy.

86

Table 4.7: Johansen Multivariate Co integration test for Modell

.-

Trace

Amax

Null

Alternative

Unadjusted

95%C.V

78.658**
43.113**
23,978
17.941
10.128
3.0392
2.3835

46.231 **
40.078**
33.877
27.584
21.132
14.265
3.841

k=l
r=l
r=2
r=3
r=4
r=5
r=6
r=7

r=O
r~l
r~2

r~3
r~4

r<5
r<6

Unadjusted

95%C.V

r=2
179.241 **
100.583**
57.470
33.492
15.550
5.423
2.383

125.615
95.754
69.819
47.856
29.797
15.495
3.841

Notes: k denotes the lag length. Asterisk (**) indicate that it is statically significant at 5 percent level

Modell:

Estimated Modell:

LGDP= 0.7644 + O.1930LM2,- O.0026LFD,- O.0792LER,- O.1738LIPt +

(0.0011) (0.0000)

(0.8565)

(0.1121)

(0.0023)

O.5488LHCE, + 0.3004LXt+ Et
(0.0000)

(0.0000)

R2 = 0.9997
Note: P-Values are given in parentheses.

From the p-value of the estimated Modell, the variables that are significant
are LM2, LIP, LHCE and LX. All the p-value for these variables are smaller than 10
percent significant level. This suggests that these variables are important detenninant
ofLGDP. as in order to be significance, the p-value must be smaller than 10 percent,
5 percent and 1 percent significance level. So, in Model will include LM2, LX, LIP
and LHCE as the independent variables.
87

4.3.2 Model 2

Similar to Modell, Model 2 is also checked whether the variables are

cointegrated with each other before proceeding to model estimation. As shown in
Table 4.8, for the null hypothesis of r = 0, the trace test is 93.586 and it is larger than
the 5 percent critical value of 69.819 thus we can reject the null hypothesis. On the
other hand, for the same null hypothesis the test statistic for max eigenvalue is
39.977 is also larger than its 5 percent critical value of 33.877, in which resulted in
rejection of the null hypothesis. For the null hypothesis of r :5 1 the test statistic for
trace test is 53.608 which are larger than the 5 percent critical value of 47.856. This
leads to the rejection of the null hypothesis. Meanwhile, for the same null hypothesis
the test statistic for max eigenvalue of 38.811 is also larger than its 5 percent critical
value of27.584 thus rejecting the null hypothesis. From r:5 2 onwards, the entire test
statistics for trace and max eigenvalue are smaller than their 5 percent critical values,
so the null hypothesis cannot be rejected. As results, it can be concluded that Model
2 has two co integrating vectors which proves that Model 2 is cointegrated with one
another thus allowing estimation of model 2 can be carried out.

Table 4.8: Johansen Multivariate Cointegration test for Model 2

Amax

Null

Alternative

Unadjusted

95%C.V

k=l
r=O
r:::;;l
r<2
r<3
r:::;;4

r=l
r=2
r=3
r=4
r=5

39.977**
38.811 **
9.920
2.984
1.893

33.877
27.584
21.132
14.265
3.841

Trace
Unadjusted

95%C.V

r=2
93.586**
53.608**
14.797
4.877
1.8928

69.819
47.856
29.797
15.495
3.841

Notes: k denotes the lag length. Asterisk (**) indicate that it is statically significant at 5 percent level.

88

Model 2:

Estimated Model 2:

L7Jf5p= 0.3685 + 0.1832LM2t + 0.607954LHCEt + 0.2986LX,- 0.2350LIP1 + Ct

(0.1912)

(0.0000)

(0.0000)

(0.0008)

(0.0000)

R2 = 0.9996
Note: P-Values are given in parentheses

4.3.3 Model 3

In order to obtain the variables that will be included in Model 3, bivariate

correlation will be conducted to observe the variable that has the highest correlation
value to LGDP. However, before that normality test will be carried out first to check
whether the variables are normally distributed for determining which correlation
coefficient will be used for the correlation test. The result obtained in the normality
test is shown in Table 4.9 in which it shows that all the seven variables are normally
distributed thus Pearson correlation coefficient will be used for the correlation test.

The result of the correlation test is shown in Table 4.10 and Table 4.11. As
seen from Table 4.14 and 4.15, two correlation tests is done. One is conducted for
the full sample and another is for the in sample. From both of the correlation test, a
consistent finding can be seen where both of the correlation test pointed out that
LHeE has the highest correlation value towards LGDP with 0.999. Thus, the
variables involve in Model 3 include LGDP and LHeE with LGDP as the dependent
variable and LHeE as the independent variable.
89

Table 4.9: Jarque-Bera Nonnality Test

Variables
LGDP
LER
LFD
LHCE
LIP
LM2
LX

Jarque-Bera
2.201
4.024
2.026
2.083
3.316
2.270
2.733

P-Value
0.363
0.134
0.393
0.353
0.191
0.321
0.255

Note: The null hypothesis for Jarque-Bera is normally distributed. Asterisks (U*), (**) and (*) denote statistical
significant at 1%, 5% and 10% level.

Table 4.10: Pearson Correlation Test for 1970 to 2010

Pearson Correlation
LGDP
LM2
LER
LHCE
LIP
LX
LFD

P-Value (2-Tailed)

0.998***
0.679***
0.999***
0.994***
0.997***
-0.515***

0.000
0.000
0.000
0.000
0.000
0.001

Note: Asterisk (***) denote correlation is significant at 0.01 level (2-tailed).

Table 4.11: Pearson Correlation Test for 1970 to 2008

LGDP
Pearson Correlation
LGDP
LM2
LER
LHCE
LIP
LX
LFD

1
0.998***
0.676***
0.999***
0.994***
0.997***
-0.436***

P-Value
0.000
0.000
0.000
0.000
0.000
0.006

Note: Asterisk (***) denote correlation is significant at 0.01 level (2-tailed).

Based on the correlation test also, it can be observed that LER, LHCE, LIP,
LM2 and LX have positive relationship with LGDP while LFD is negatively related
to LGDP. All of the independent variables estimated sign is in line with the expected
90

sign according to the theory stated in chapter 3 with the exception of LER and LFD
in which the expected sign should be negative as higher value of LER means
depreciation thus when the exchange rate of Malaysia appreciates means that GDP of
Malaysia will increase so in other words, as the value of LER decreases it indicate
that appreciation happen. However, the positive sign for LER in Malaysia may
happen because Malaysia is an export oriented country thus as the exchange rate of
Malaysia depreciate it indicates that other countries will buy more of Malaysia
products thus increasing the GDP. Another reason also might be that the fixing of the
Ringgit from 1998 to 2005 has causes it to deviate from the expected sign as the
theory is based on normal condition.

In addition, Pearson correlation test also indicates the impact of each variable
toward LGDP. For LM2, LHeE, LIP and LX they have a correlation coefficient of
0.998, 0.999, 0.994 and 0.997 respectively which show that they have strong impact
to LGDP. On the other hand, for LER, it has a fair impact toward LGDP with only a
correlation coefficient of 0.679 for the full sample and 0.676 for the in sample.
Similarly, for LFD, it has a rather fair impact with a correlation coefficient of 0.515
for the full sample and a weak impact with only 0.436 for the in sample. This shows
that these variables' impact to the LGDP increases as time passes. Fixed deposit has
a rather weak relationship with LGDP most probably because Malaysia does not
really alter their interest rate much at most the interest rate is only altered in between
3 percent to 4 percent. All the variables are significant towards LGDP as all of their
p-value is less than 10 percent significant leveL The only difference is that LER and
LFD do not have much impact on LGDP compared to the others.

91

Table 4.12: Johansen Multivariate Cointegration test for Model 3

Null

Alternative

Amax
---------

Unadjusted

95% C.V

k=l
r=O
r<l

r=l
r=2

Trace
Unadjusted

.~-----------------

95%C.V

r=l

27.011
15.892
32.725**
20.262
5.714.~~~~~~-=~~~~------~~~~----9.165
5.714
9.165

Notes: k denotes the lag length. Asterisk (**) indicate that it is statically significant at 5 percent level.

Model 3 is also checked whether the variables are cointegrated with each other
before proceeding to model estimation. As shown in Table 4.12, for the null
hypothesis of r = 0, the trace test is 32.725 and it is larger than the 5 percent critical
value of 20.262. Thus, we can reject the null hypothesis. On the other hand, for the
same null hypothesis the test statistic for max eigenvalue is 27.011 and it is also
larger than its 5 percent critical value of 15.892 thus, resulted in rejection of the null
hypothesis. For r S 1, test statistic for trace is 5.714 which is lower than the 5 percent
critical value of 9.165 thus resulted in not rejecting the null hypothesis, max
eigenvalue with 5.714 is also smaller than its 5 percent critical level of9.165 thus the
null hypothesis also cannot be rejected. As a result, it can be concluded that model 3
has one cointegrating vector which means the variables in model 3 are cointegrated
with one another showing that long run relationship exist.

The result of the estimation for Model 3 can be seen below. After aU the model
is estimated the performance of the three models will be evaluated which will be
discussed in the next section.

92

Model 3:

LGDPt = /lo+ /l,LHCEt + E t

Estimated Model 3:
LGDP= -0.1645 + 1. o796LHCEt + Et
(0.1 096)

(0.0000)

R2 = 0.9986
Note: P-Values are given in parentheses

4.3.4 Performance of the Fundamental models

Table 4.13: Forecasting Performance of Fundamental models

Full sample

Out-samEle

RMSE

MAPE

In-satnEle
RMSE

Modell

0.0239

0.1645

0.0216

0.1539

0.0512

0.3713

Model 2
Model 3

0.0265
0.0533

0.1870
0.3464

0.0232
0.0447

0.1734
0.3116

0.0627
0.1390

0.4515
1.0250

Models

MAPE

RMSE

MAPE

Notes: Full sample eover the range from 1970 to 2010, In-sample from 1970 to 2008 and Out-sample from 2009
to 2010. The estimation period is from 1970 to 2008 while the foreeasting period from 2009 to 2010.

From Table 4.13, it can be seen that Model 1 forecasts the LGDP the best
compare to Model 2 and Model 3, based on RMSE and MAPE values for all full
sample, in-sample and out-of sample. This leads to the fact that the accuracy to
forecast the LGDP increases as more variables that could affect the LGDP is
included. Thus, among the three fundamental models present, Model 1 is chosen as
the best model to represent the fundamental model. This finding, however, is in
sharp contrast to the principle of parsimony. It would always be desirable if one
could predict LGDP with the least numbers of macroeconomic variables.
93

4.4

Random Walk Model

The final model that will be use to forecast the LGDP is random walk model or
also known as naive model which had been used by a lot of people to forecast certain
variables (Agwuegbo, Adewole & Maduegbuna, 2010; Pesaran & Oick, 2008) The
model for the purpose of this study is computed as:

LGDPt

= LGDPt _ +
1

where t is the year estimated and t - 1 is a year before the estimated year.

is the

average change between period. Random walk model will be label as R WM (1). The
forecast accuracy of random walk model is evaluated in Section 4.4.1.

4.4.1 Forecast Accuracy for Random Walk Model

Table 4.14: Forecasting Performance of Random Walk Model

.......c...Fu.;;..I_I-'--sam.;;....:...PLI..:...e_ _ _ _ _I___._n-___._s-'-'--am.L. pl.. :.e_ _ _ _ _ _o.::....:::..ut.;;..-s...:.ac--mple
Models
RWM(l)

.-

RMSE

MAPE

RMSE

MAPE

RMSE

MAPE

0.1272

1.0209

0.1283

1.0344

0.1042

0.7640

Notes: Full sample cover the range from 1970 to 2010, In-sample from 1970 to 2008 and Out-sample from 2009
to 2010. The estimation period is from 1970 to 2008 while the forecasting period from 2009 to 2010.

Table 4.14 show the forecast accuracy for random walk model for full sample,
in sample and out-of sample. From the table, it can be observed that the accuracy of
forecasting using a random walk model is quite low in which all the RMSE for full
sample, in sample and out sample are higher than 10 percent. The next section will
compare ARlMA model, fundamental models and random walk model to identify
the models which could provide the most accurate forecast.

94

4.5 Comparison between ARIMA, Fundamental models and Random Walk

Model

Table 4.15: Forecasting Performance of ARIMA, Fundamental models and Random

Walk Models
Models

FuJI

RMSE

MAPE

RMSE

MAPE

RMSE

MAPE

ARlMA(4,1,3l 0.0854

0.6007

0.0786

0.5657

0.1416

1.0492

Model 1
RWM (1)

0.0239

0.1645

0.0216

0.1539

0.0512

0.3713

0.1272

1.0209

0.1283

1.0344

0.1042

0.7640

Notes: Full sample cover the range from 1970 to 2010, In-sample from 1970 to 2008 and Out-sample from 2009
to 20 10. The estimation period is from 1970 to 2008 while the forecasting period from 2009 to 20 I O.

Table 4.15 above shows the forecasting performance for selected best model
ARIMA, the best fundamental models and the random walk model. It can be
concluded that fundamental models could predict the LGDP more accurately than
ARlMA and random walk model which is proven by the lower RMSE and MAPE of
the full sample, in sample and out sample of the fundamental models. Moreover, the
high value of RMSE and MAPE seen for the random walk model indicate that
LGDP of Malaysia cannot be forecast accurately by only using the LGDP of the
previous years. Out of the three models, Model 1 which included all the independent
variables could predict the future GDP the best compare to the others. This is
followed by the ARlMA (4,1,3) model. The random walk model performed the
worst in this study.

9 This is the best model for ARlMA from the other 24 models as it has the lowest RMSE and MAPE values for
full sample, in sample and out of sample.

95

---------Chapter Five
Conclusion and Recommendation
5.0

Introduction

The final chapter in this study will discuss on the finding that was obtained
from the previous chapter together with its interpretation and also some policy
recommendation that could help the country. Moreover, the limitation of this study
and recommendation for further studies are also discussed. The last part of this paper
is conclusion which will concludes the whole study.

This study estimates the Autoregressive Integrated Moving Average (ARIMA)

time series models and fundamental models for the logarithm form of the GDP of
Malaysia for the period 1970 to 2008. This is in-sample period while the out-of
sample period for forecasting purpose ranges from 2009 to 2010. Before these
models are estimated, unit root tests had to be carried out to ensure that the variables
are stationary especially for the fundamental models in which the variables must be
integrated at order one, 1(1) to enable execution of Johansen cointegration test. If the
variable are l(l), they must be cointegrated before they can be estimated in OLS. The
full samples, in-samples and out-of samples forecast accuracy resulting from all the
three types of model is then compared. The major findings obtained in the study are
discussed in detail in the next section.

96

5.1

Summary of Findings

In forecasting a country GDP, there are lots of methods that can be applied and
each method has their own forecast performance in which one model might have
high forecast accuracy for a particular country and there might also be possibility of
having low forecast accuracy for the same model but for different countries.

Based on the Phillip-Perron (PP) test, it can be concluded that logarithm of

gross domestic product (LGDP), logarithm of broad money (LM2), logarithm of
fixed deposit (LFD), logarithm of industrial production (LIP), logarithm of exchange
rate (LER), logarithm of export (LX) and logarithm of household consumption
expenditure (LHCE) are all integrated at order one, 1(1). This leads to the use of
ARIMA model instead of ARMA model and also directly qualified the variables to
proceed to Johansen cointegration test. Three fundamental models are estimated in
this study. The first model consists of all the variables which are LGDP, LM2, LER,
LFD, LIP, LX and LHCE. The second model consist only the variables that are
significant from the full models which are LGDP, LM2, LHCE, LX and LIP because
LFD and LER. Lastly, the third model only estimates LGDP towards the variable
with the highest correlation in which is LHCE with 0.999 which indicates a very
strong positive linear relationship. Bivariate correlation is used to check for the level
of correlation and Pearson coefficient is used when testing the correlation because all
the variables are normally distributed. All the variables used in the three models are
cointegrated.

From Pearson correlation test, for LM2, LHCE, LIP and LX they have a
correlation coefficient of 0.998, 0.999, 0.994 and 0.997 respectively which show that
97

they have strong impact to LGDP. On the other hand, for LER, it has a fair impact
toward LGDP with only 0.679 for the full sample and 0.676 for the in sample. Then,
for LFO, it has a rather fair impact with 0.515 for the full sample and a weak impact
with only 0.436 for the in sample. This indicates that LM2, LHCE, LIP and LX
could be used to stimulate the LGOP effectively.

There are 24 models involve each for ARIMA models with and without
constant. Based on Root Mean Square Error (RMSE) and Mean Absolute Percentage
Error (MAPE), ARIMA models with constant performed better than ARIMA model
without constant. Consistently, based on RMSE and MAPE also, ARIMA (4,1,3)
model is chosen as the best model from the 24 model in ARIMA with constant
models. ARIMA (4,1,3) is chosen because it has the lowest RMSE and MAPE value
for full sample, in-sample and out-of sample. On the other hand, model 1 from the
three fundamental models is chosen as the best model as a result of it smallest values
in term ofRMSE and MAPE for full sample, in-sample and out-of sample.

As a result, out of the three models compared, fundamental model had the
lowest RMSE and MAPE in term of full sample, in-sample and out-of sample which
makes it to be the nest model among ARIMA (4,1,3) model and random walk model.
The second best among them is ARIMA (4,1,3) and the model which forecast the
worst among them is the random walk model.

According to Gilbert (1995), a relatively simple model specification should be

aim in other words the simpler the model the better it will be. However, this
statement is in contrast with the results obtained from the study, in which it can be
seen that for forecasting GOP in Malaysia in order to get more accurate result, more
98

variables is needed. Nevertheless, seeing this model is the best, it is worth to take
some time to model it.

5.2

Policy Implication and Suggestion

Economic stability is given much attention these days especially recently

because of political unrest that happen in the Middle East. It causes the economic
stability of the Middle East to be disrupted. A stable economy actually portrays a
positive image and an excellent economic positioning in other words strong
economic stability can attract other countries to come to our countries to invest.
Because of this, the factors in detennining economic stability should be given
attention. The detenninant factors such as exports, consumption expenditure and
exchange rate can be a threat if it is not managed well. Especially for Malaysia as
seeing from the correlation test, aside from exchange rate which give a rather fair
impact to GDP and also interest rate which measure in the fonn of fixed deposits that
give a weak impact, the other four factors studied which include the money supply,
consumption expenditure, exports and industrial production could have strong
impact to the economy of Malaysia as they have a strong and significant relationship
with GDP. This is in line with Kogid, et. al (2010) results that state consumption
expenditure and export play important roles as detenninant factors to a country's
economic growth. Thus the policies in Malaysia should focus on these four aspects
in order to ensure a higher GDP for the country.

99

5.2.1 Public Expenditure

Household consumption expenditure is an important detenninant of Malaysia

GDP. One of the ways to increase the households' consumption expenditure in
Malaysia in order to increase the GDP is by increasing public expenditure. Public
expenditure in this sense means government social expenditures on education, health
and pension. A study in China shows that declining household consumption is due to
the structural changes in economy and also the reduced provision of support for
education, health and old age pension by the government. Thus, in order to increase
household consumption expenditure three channels had been identified that could
help and these channels include direct income channel, distributional channels and
also insurance channel and this is proven by the household survey data that shows
the increase in public spending for these three categories could be significant
(Baldacci et aI., 2010).

An example of government public expenditure in health is that when the

charges for health care reduces thus this means that people would have more money
to consume instead of saving it as fear that someday they will fall seriously ill and
did not have the money to pay for their medical bill. Moreover, Barnett and Brooks
(2010) also state that the increased efficiency in the delivery of public health care
will be a key factor in impacting consumption behaviour. All these are trying to
show that spending on education, healthcare and the pension of the retire worker
could increase household consumption and increases the country's GDP.

100

5.2.2 Foreign Market Access

Export is another major determinant for the performance of GDP in Malaysia.

Export can be increased by improving foreign market access. Foreign market access
can be improved by further lowering the trade barriers at all stages of the
development of external sector which in term is actually attracting foreign investors
to come and invest in Malaysia. By facilitating international market access, it would
also be able to facilitate trade especially for intermediate goods thus increasing the
returns of our export.

This in term proven by Fugazza (2004), in which their result suggest that the
structure of external sector and foreign market access are complementing each other
at low levels performance however these two variable become substitutes at higher
performance level which means that as external sector develops, the access towards
international market become more significant in explaining the performance of
exports. However, focussing on better international market access only and
neglecting the supply condition is an unwise choice as it unproductive towards our
country exports which similar to the concept of trying to accept a lot of order but do
not have a sufficient supplies to provide all the goods that the customers have order

in which could jeopardize the name of the company and also the trust of the
customers and this goes the same to a country.

101

5.2.3 Monetary Policy

Monetary policy affects spending and the supply of money by altering the
interest rate. Thus, monetary policy in this case could help in eliminating the cost of
money illusion by correcting their distortions on consumption and savings decision
and this is done by choosing a growth rate of the money supply. Money illusion
happens when people are confused between nominal and real magnitudes. Thus, this
monetary policy could implement a specific nonzero expected inflation rate or a
constant nominal interest rate so that the distortions from misperception can offset
each other and thus increases spending of the people. However, this situation also
depends on the existence of risk averters, if there are a lot of risk averters presence
then they will tend to avoid it at all cost even though the policy is helping them
(Miao and Xie, 2007).

As compared to fiscal policy such as taxes on labour income would also

influence GDP and provide different effect and according to Engen and Skinner
(1992), there is evidence on the empirical record that point out the importance of
fiscal policy towards output growth. However, Ajayi (1974) result found that the
beta coefficients that changes in monetary action were greater than fiscal action. As
Nigeria is similar to Malaysia in term that both these countries are developing
countries thus most of the time emphasis is always given on monetary policy rather
than fiscal policy (Ajisafe and Folorunso, 2002).

102

5.3

Limitations

One of the limitations identified in this study is the unavailability of data for
some variables as there are some variables that identified useful by others for
example Foreign Direct Investment (FDI), Government 5 years bond, share price and
etc. data is unavailable hence these variables are to not be taken into considerations.

Another limitation is the less observations of the out-of sample which is only
two years thus for more reliable result, more years or observations should be
included in the out-of sample forecast.

5.4

Recommendations for future studies

A field for future research would be to examine the model performance at out
of sample forecast horizon longer than 2 years in order to observe the differences.
Moreover, as the accuracy for forecasting the GDP of Malaysia increases with more
variables then future studies should extend the forecasting to other essential
macroeconomIC variables, such as inflation rate and considering more complex
models.

Another way to further the research is to suggest more countries to see whether
the model in the study could also performed well for other countries. In addition,
further studies can also be conducted to formulate computerize programming for the
generation of future GDP forecast.

103

5.5

Contribution of Study

This study helps to identify important factors in order to enhance GDP growth.
Moreover, various forecasting techniques in forecasting the GDP had also been
conducted and it is found out that fundamental models performed better than time
series and random walk model thus more time should be spent to model an accurate
model in forecasting the GDP by the government.

5.6

Concluding Remark

According to this study, broad money, household consumption expenditure,

industrial production and exports are very important to the growth of Malaysia
economy. Thus, the government of Malaysia should emphasize more on sectors and
policy involving with these factors in order to promote better economic growth of
the country. For example, government could increase social public expenditure in
term spending more on developing health care, education and create a better policy
in the pension system for the retired so that the people would be less worried about
their income or money as if they worried less than they tend to consume more thus
increasing household consumption and increasing the GDP as the effect of the strong
positive relationship between both of this variables.

Consumption expenditure could also be increased by providing better

monetary policy in other words alter the interest rate as such that they offset the cost
of money illusion so that people would spend more. Moreover, government could
focus on foreign market access by taking into consideration of the supply condition

104

in order to attract more investors and trading partners to boost the country's export
and increases the country's LGDP.

In evaluating the forecasting accuracy, root mean square error (RMSE) and
mean absolute percentage error (MAPE) had been used and the models that been
used to forecast the LGDP of Malaysia are ARIMA model, 3 models for the
fundamental models and one random walk model. All these models are then
compared based on RMSE and MAPE. From the comparison, it is found that
fundamental model could forecast the LGDP of Malaysia with the highest accuracy.
This is in contrast with Gilbert (1995) that states simple model specification is better
as from the results of this study it is seen that as more variables in connection with
the LGDP are included, the accuracy of the model will increase. Lastly, it is noted
that the objectives of examining the determinant of GDP, forecasting the GDP and
evaluating the forecasting accuracy of the forecast for Malaysia are achieved in this
study.

105

References

Acquah, H. D. G. (2010). Comparison of Akaike Infonnation Criterion (AIC) and

Bayesian Infonnation Criterion (BIC) in Selection of an Asymmetric Price
Relationship. Journal of Development and Agricultural Economiecs. 2(1),
001-006.
Agwuegbo, S. O. N., Adewole, A. P. & Maduegbuna, A. N. (2010). A Random
Walk Model for Stock Market Prices. Journal ofMathematics and Statistics.
6(3), 342-246.
Ajisafe, R. A. & Folonrunso, B. A. (2002). The Relative Effectiveness of Fiscal and
Monetary Policy in Macroeconomic Management in Nigeria. The African
Economic and Business Review. 3(1), 23-40.
Amanah Mutual Berhad. (2009). GDP. Retrieved October 23, 2010, from
http://www.ambmutual.com.my/publicationsieconomic%20reports/2009/May
II Q2009%20GDP%20-%20a%20weake~1020than%20expected%20start.pdf
Anaman, K. A. (2004). Detenninants of Economic Growth in Brunei Darussalam.
Journal ofAsian Economics, 15, 777-796.
Andersson, J. (2007). Forecasting Swedish GDP Growth. Unpublished Master's
thesis, Lund University, Lund, Sweden.
Angelini, E., Banbura, M. & Runstler, G. (2008). Estimating and Forecasting the
Euro Area Monthly National Accounts from A Dynamic Factor Model.
European Central Bank Working Paper, No. 953.
Angelini, E., Camba-Mendez, G., Giannone, D., Runstler, G. & Reichlin, L. (2008).
Short-term Forecasts of Euro Area GDP Growth. European Central Bank,
Working Paper No. 949.
Ansari, M. 1. (2002). Impact of Financial Development, Money, and Public Spending
on Malaysian National Income: An Econometric Study. Journal of Asian
Economics, 13, 72-93.
Arnold, R. A. (2008). Economics. USA: South-Western Cengage Learning.
Awokuse, T. (2002). Is the Export-Lead Growth Hypothesis valid for Canada.
Department of Food and Resources Economics Staff Paper, No. 15823.
Bai, J., & Ng, S. (2002). Detennining the Number of Factors in Approximate Factor
Models, Econometrica, 70, 191-221.

106

Baldacci, E., Callegari, G., Coady, D., Ding, G., Kumar, M., Tommasino, P., & Woo
J. J. (2010). Public Expenditures on Social Programs and Household
Consumption in China. International Monetary Fund Working Paper, No.
10/69.
BaneIjee, P. K., & Adhikary, B. K. (2009). Dynamic Effects of Interest Rate and
Exchange Rate Changes on Stock Markets Returns in Bangladesh. Journal
ofSocial and Management Sciences, 38(1).
Barnett, S. & Brooks, R. (2010). China: Does Government Health and Education
Spending Boost Consumption. International Monetary Fund Working Paper,
No. 10/16.
Bartov, E., & Bodnar, G.M. (1994). Firm valuation, earnings expectations,
exchange-rate exposure effect. Journal ofFinance, 49, 1755-1785.
Beechey, M., & Osterholm, P. (2010). Forecasting Inflation in an Inflation-Targeting
Regime: A Role for Informative Steady-State Priors. International Journal of
Forecasting, 26, 248-264.
Bidarkota, P. V. (1998). The Comparative Forecast Performance of Univariate and
Multivariate Models: An Application to Real Interest Rate Forecasting.
International Journal ofForecasting, 14, 457-468.
Bolboaca, S. D., & Jantschi, L. (2006). Pearson versus Spearman, Kendall's Tau
Correlation Analysis on Structure-Activity Relationships of Biologic Active
Compounds. Leonardo Journal ofSciences, 19, 179-200.
Box, G., & Jenkins, G. (1984). Time-Series Analysis: Forecasting and Control. San
Francisco, CA: Holden Day.
Brand, C., Reimers, H. E. & Seits, F. (2003). Forecasting Real GDP: What role for
Narrow Money. European Central Bank Working Paper, No. 254.
Breusch, T. S. (1978). Testing for Autocorrelation in Dynamic Linear Models.
Australian Economic Paper, 17, 334-355.
Bushman, M. (2007). Why is Forecasting Important to an Organization. Retrieved
October
31,
2010
from
http://www.associatedcontent.com/article1200360/why_is_forecasting_impor
tant to an.html?cat=3
Byers, S. L. & Nowman, K. W. (1998). Forecasting U.K. and U.S. Interest rates
using Continuos Time Term Structure Models. International Review of
Financial Analysis, 7(3), 191-206.
Caraiani, P. (2008). Forecasting Romanian GDP using a Small DSGE modeL
Romanian Journal ofEconomic Forecasting, 9(1), 182-191.
107

JiML

Carriero, A., Kapetanios, G. & Marcellino, M. (2009). Forecasting Exchange Rates

with a large Bayesian VAR. International Journal of Forecasting, 25, 400
417.
Cavanaugh, J. E., & Neath, A. A. (1999). Generalizing the Derivation of the
Schwarz Information Criterion. Communications in Statistics-Theory and
Methods, 28(1), 49-66.
Central Intelligence Agency (CIA). (2010). Malaysia. Retrieved October 23, 2010,
from
https:llwww.cia.govllibrary/publications/the-world
factbook/geos/my.html
Charette, D. E. (2005). Malaysia in the Global Economy: Crisis, recovery, and the
Road Ahead. Boston: University of Massachusetts Boston.
Chen, B. & Feng, Y. (1996). Some Political Determinants of Economic Growth:
Theory and Empirical Implications. European Journal ofPolitical Economy,
12, 609-627.
Cheng, M. Y. & Tan, H. B. (2002). Inflation in Malaysia. International Journal of
Social Economics, 29(5), 411-425.
Choi, I. (1992). Effects of Data Aggregation on the Power of Tests for a Unit Root.
Economics Letters, 40, 397-401.
Chong, L. L. & Tan, H. B. (2008). Exchange Rate Risk and Macroeconomic
Fundamentals: Evidence from Four Neighbouring Southeast Asian
Economies. International Research Journal of Finance and Economics, 16.
1450-2887.
D'Agostino, A., McQuinn, K. & O'Brien, D. (2008). Now-Casting Irish GDP.
Central Bank & Financial Services Authority of Ireland Research Technical
Paper, No.9.
Dickey, D. A. & Fuller, W. A. (1979). Distribution of the Estimators for
Autoregressive Time Series with a Unit Root. Journal of the American
Statistical Association, 74 (366),427-431.
Drabble, 1. H. (2010). Economic History of Malaysia. Retrieved October, 24, 2010
from http://eh.netlencyclopedialarticle/drabble.malaysia
Dritsaki, M., Dritsaki, C. & Adamopoulos. (2004). Causal Relationship between
Trade, Foreign Direct Investment and Economy Growth for Greece.
American Journal ofApplied Sciences, 1 (3), 230-235.
EconomyWatch. (2010). Malaysia Finance, Finance Malaysia. Retrieved October
24,
2010,
from
http://www.economywatch.comlfinancelMalaysia
finance.html
108

"'4

Encyclopedia of the Nations. (2010). Malaysia-Overview of Economy. Retrieved

October,
24,
2010
from
http://www.nationsencyclopedia.com/economies/Asia-and-the
PacificlMalaysia-OVERVIEW-OF -ECONOMY.html
Engen, E. M. & Skinner, J. (1992). Fiscal Policy and Economic Growth. National
Bureau of Economic Research Working Paper, No. 4223.
Fama, E. F. (1981). Stock Returns, Real Activity, Inflation, and Money. The
American Economic Review, 71(4), 545-565.
Fama, E. F. (1990). Stock Returns, Expected Returns, and Real Activity. The
Journal ofFinance, 45(4), 1089-1108.
FoIster, S. & Henrekson, M (1999). Growth in the public Sector: A critique of the
Critics. Research Institute of Industrial Economics Working Paper Series,
No. 492.
Fernandez, V. (2006). External dependence in European capital markets. Journal of
Applied Economics, 9(2), 275~293.
Fletcher, D. J. & Gulley, O. D. (1996). Forecasting the Real Interest Rate. North
American Journal ofEconomics & Finance, 7(1), 55-76.
Franck, P. & Young, A. (1972). Stock price reaction of multinational firms to
exchange realignments. Financial Management, 1,66-73.
Fugazza, M. (2004). Export Performance and its Determinants: Supply and Demand
Constraints. Policy Issues in International Trade and Commodities Study
Series, No. 26.
Giannone, D., Reichlin, L. & Small, D. (2008). Nowcasting: The Real-Time
Informational Content of Macroeconomic Data Journal of Monetary
Economics, 55(4), 665-676.
Gilbert, P. D. (1995). Combining V AR Estimation and State Space Model Reduction
for Simple Good Predictions. Journal ofForecasting, 14, 229-250.
Godfrey, L. G. (1978). Testing Against General Autoregressive and Moving Average
Error Models when Regressor includes Lagged Dependent Variables.
Econometrica, 46, 1293-1302.
Gounder, R. (1999). The Political Economy of Development: Empirical Evidence
from Fiji. Economic Analysis & Policy, 29(2), 133-150
Grossmann, A. & McMillan, D. G. (2010). Journal of International Financial
Markets, Institutions & Money. Journal of International Financial Markets,
Institutions & Money, 20, 436-450.
109

Guegan, D. & Rakotomarolahy, P. (2010). Alternative Methods for Forecasting

GDP. Centre d'Economie de la Sorbonne Working Paper, No. 10065.
Gupta, R. (2006). Forecasting the South African Economy with VARs and VECMs.
South Ajrican Journal ofEconomics, 74(4),611-628.
Hansson, J., Jansson, P. & Lof, M. (2005). Business Survey Data: Do they help in
forecasting GDP Growth. International Journal ofForecasting, 21, 377-389.
Hoehn, J. G., Gruben, W. C. & Fomby, T. B. (1984). Some Time Series Methods of
Forecasting the Texas Economy. Federal Reserve Bank of Dallas Research
Paper, No. 8402.
Jarque, C. M. & Bera, A. K. (1987). A Test for Normality of Observations and
Regression Residuals. International Statistical Review, 55, 163-172.
Johansen, S. & Juselius, K. (1990). Maximum Likelihood Estimation and Inference
on Cointegration with Applications to the Demand for Money. Oxford
Bulletin ofEconomics and Statistic, 52(2), 0305-9049.
Keynes, J. M. (1936). The General Theory ofEmployment, Interest and Money. New
York: Harcourt, Brace and Company.
Kogid, M., Mulok, D., Lim, F. Y. & Mansur, K. (2010). Determinant Factors of
Economic Growth in Malysia: Multivariate Co integration and Causality
Analysis. European Journal of Economics, Finance and Administrative
Sciences, 24, 1450-2275.
Krajewski, J. (2009). Estimating and Forecasting GDP in Poland with Dynamic
Factor Model. Dynamic Econometric Models, 9.
Kweka, C. & Morrissey, O. (1998). Impacts of Economic Growth on Consumption
Expenditures, the Case of Tanzania. Tanzanian Economic Journal, 15(1), 1
34
Liu, G. L., Gupta, R. & Schaling, E. (2010). Forecasting the South African
Economy: a Hybrid-DSGE Approach. Journal of Economic Studies, 37(2),
181-195.
Lu, Y. (2009). Modelling and Forecasting China's GDP data with Time Series
Models. Unpublished thesis, Hogskolan Dalarna, Sweden.
Malaysia Canada Business CounciL (n.d.). Malaysia's Economic Profile. Retrieved
October 24, 2010, from http://www.malaysia-canada.comlinfo-malaysia
economy.htm

110

Mittnik, S. & Zadrozny, P. (2004). Forecasting Quarterly German GDP at Monthly

Intervals using Monthly IFO Business Conditions Data. Center for Economic
Studies and IFO Institute for Economic Research (CESIFO) Working Paper,
No. 1203.
Mehrothra, A. & Sanchez-Fung, J. R. (2008). Forecasting Inflation in China. Bank
of Finland Institute for Economies in Transition (BOFIT) Discussion Paper,
No.2.
Miao, J. J. & Xie, D. Y. (2007). Monetary Policy and Economic Growth under
Money Illusion. Boston University Department of Economics WWorking
Paper Series, No. Wp2007-045.
Mok, H. M. K. (1993). Causality of Interets Rate, Exchange Rate and Stock Prices at
Stock Market Open and Close in Hong Kong. Asia Pasific Journal of

Management, 10(2), 123-143.

Nieh, C.C. & Lee, C.F. (2001). Dynamic relationship between stock prices and
exchange rates for G-7 countries. Quarterly Review of Economics and
Finance, 41 (4), 477-490.
Obamuyi, T. M. (2009). An Investigation of the Relationship between interest rates
and economic growth in Nigeria, 1970-2006. Journal of Economics and
International Finance, 1(4), 093-098.
Onour, I. (2007). Testing Efficiency Performance of an Underdeveloped Stock
Market. Munich Personal Repec Archive Paper, No. 15020.
Patinkin, D. (1965). Money, Interest and Prices. New York: Harper & Row.
Pesaran M. H. & Pick, A. (2008). Forecasting Random Walks under Drift Instability.
Cambridge Working Papers in Economics, No. 0814.
Phillips, P. C. B. & Perron, P. (1988). Testing for a Unit Root in Time Series
Regression. Biometrika, 75(2), 335-346.
Pigou, A. C. (1943). The Classical Stationary State. The Economic Journal, 53, 343
351.
Rapach, D. E., & Weber, C. E. (2004). Are Real Interest Rates really nonstationary?
New Evidence from Tests with Good Size and Power. Journal of
Macroeconomics, 26, 409-430.
Runstler, G. & Sedillot, F. (2003). Short-term Estimates ofEuro Area Real GDP by
Means of Monthly Data. European Central Bank Working Paper Series, No.
276.

111

..

Saraogi, R. (2008). GDP Forecast for Australia. Munich Personal RePEc Archive
(MPRA) Paper, No. 22797.
Schneider, M. & Spitzer, M. (2004). Forecasting Austrian GDP using the
Generalized Dynamic Factor Model. Oesterreichische Nationalbank Working
Paper, No. 89.
Sentsho, J. (2000). Export Revenues as Determinants ofEconomic Growth: Evidence
from Botswana. Unpublished Thesis, University of Botswana, Gaborone,
Botswana.
Serrato, J. C. S. (2006). Forecasting Mexico's Inflation: the Effects of an Inflation
Targeting Regime. Proceedings of the 5 th Annual Carroll Round 2,
Washington, DC.
Sims, C. A. (1980). Macroeconomics and Reality. Econometrica, 48, 1-48.
Taylor, J. B. (2007). Principles of Macroeconomics. Boston: Houghton Mifflin
Company.
U.S. Department of State. (2010). Background Note: Malaysia .. Retrieved October
23,2010, from http://www.state.gov/r/pa/eilbgn/2777.htm
Weber, A. A. (2009). The role of Forecasting for Central Banks. Conference on
Forecasting and Monetary Policy, Berlin, Germany.
Wei, S. G., & Awazu, Y. (2008). Vagaries of the Euro: an Introduction to ARIMA
Modeling. Case Studies in Business, Industry and Government Statistic, 2(1).
45-55.
White, H. (1980). A Heteroscedasticity Consistent Covariance Matrix Estimator and
a Direct Test of Heteroscedasticity. Econometrica, 48, 817-818.

112

Appendix A

Table AI: PP results for results for 1991QI-201OQ4

PP

Level

Variables
None

Intercept

LGDP

First Difference
Intercept & Trend

None

Finding
Intercept

3.250 (10)*

1(1)

Notes: LGDP= natural log of Gross Domestic Product. LM2= natural log of Broad money, LFD= natural log for Fixed Deposits, LER== natural log of exchange rate, LX= natural log of
exports, LIP= natural log of industrial production, LHCE= natural log of household consumption expenditure. Asterisks (""), (*"') and (*) denote statistical significant at 1%, 5% and 10%
level. Figures in the bracket 0 are the optimal lag length. The period of estimation is from 1991Q I to 20 IOQ4.

Table A2: PP results for results for 1991 Q 1-2008Q4

PP

Variables

Level
None

LGDP

Intercept

First Difference
Intercept & Trend
3.416 (5)*

None

Finding
Intercept
1(0)

Notes: LGDP= natural log of Gross Domestic Product, LM2= natural log of Broad money, LFD= natural log for Fixed Deposits, LER= natural log of exchange rate, LX= natural log of
exports, LIP= natural log of industrial production, LHCE= natural log of household consumption expenditure. Asterisks (U.), ( .... ) and (.) denote statistical significant at 1%.5% and 10%
level. Figures in the bracket () are the optimal lag length. The period of estimation is from 1991 Q I to 2008Q4.

. JtJU ,,'

Appendix B

Table Bl: Forecasting Perfonnance of ARMA models with constant (Quarterly)

ARMA

(0,0) with constant

(0,1) with constant
(0,2) with constant
(0,3) with constant
(0,4) with constant
(1,0) with constant
(1,1) with constant
(l,2) with constant
(1,3) with constant
(1,4) with constant
(2,0) with constant
(2,1) with constant
(2,2) with constant
(2,3) with constant
(2,4) with constant
(3,0) with constant
(3,1) with constant
(3,2) with constant
(3,3) with constant
(3,4) with constant
(4,0) with constant
(4,1) with constant
(4,2) with constant
(4,3) with constant
(4,4) with constant

Full sample
RMSE
MAPE
0.5292
3.9216
0.5132
3.8010
0.5047
3.7500
0.4989
3.7089
0.4937
3.6813
0.0757
0.5257
0.0849
0.5829
0.0716
0.5320
0.0617
0.4424
0.0689
0.5049
0.0819
0.5696
0.0883
0.6105
0.0706
0.5179
0.0656
0.4749
0.0673
0.4935
0.0725
0.5259
0.0722
0.5237
0.0641
0.4559
0.0628
0.4359
0.0712
0.5245
0.0726
0.5302
0.0720
0.5253
0.0694
0.5004
0.0603
0.4289
0.0654
0.4448

In-sample
RMSE
MAPE
0.4870
3.6124
0.4685
3.4447
0.4611
3.4026
0.4547
3.3603
0.4526
3.3505
0.0776
0.5454
0.0857
0.5838
0.0661
0.4974
0.0585
0.4223
0.0641
0.4743
0.0844
0.5956
0.0911
0.6314
0.0684
0.5028
0.0641
0.4664
0.0641
0.4732
0.0654
0.4812
0.0655
0.4810
0.0637
0.4556
0.0637
0.4446
0.0663
0.4930
0.0666
0.4911
0.0666
0.4898
0.0661
0.4784
0.0603
0.4326
0.0669
0.4542

Out-sample
RMSE
MAPE
0.8158
6.7039
0.7776
6.2307
0.7494
5.9391
0.7337
5.7508
0.6641
4.9030
0.0729
0.4775
0.0579
0.4006
0.1121
0.8889
0.0578
0.4549
0.0965
0.7533
0.0713
0.4570
0.0580
0.3629
0.1004
0.7948
0.0695
0.5326
0.0866
0.6786
0.1188
0.9250
0.1171
0.9083
0.0766
0.5304
0.0653
0.4388
0.0985
0.7572
0.1145
0.8819
0.1113
0.8588
0.0979
0.7101
0.0746
0.5136
0.0540
0.3722

Notes: Full sample cover the range from 1970 to 2010, In-sample from 1970 to 2008 and Out-sample from 2009
to 2010. The estimation period is from 1991 to 2008 while the forecasting period from 2009 to 2010.

...