The Determinant of Reserves in Malaysia

1.0  INTRODUCTION

This study attempts to empirically examine the determinants of reserves in Malaysia. Times series monthly data spanning from 1994 to 2013 were utilized in the study. The ADF unit root test results prove that all the variables are stationary after being differenced once. The co-integration test results further show that international reserves and the specified determinants are co-integrated. Total reserve can be defined as total reserves comprise holdings of monetary gold, special drawing rights, reserves of IMF members held by the IMF, and holdings of foreign exchange under the control of monetary authorities.^[1]The definition of export is to take or cause to be taken out of Malaysia any items by land, sea or air, or to place any items in a conveyance for the purpose of such items being taken out of Malaysia by land, sea or air or to transmit technology by any means to a destination outside Malaysia, and includes any oral or visual transmission of technology by a communication device where the technology is contained in a document the relevant part which is read out, described or otherwise displayed over the communications device in such a way as to achieve similar result (“Intangible Technology Transfer”)^[2]. Meanwhile, the definition of Import is a good or service brought into one country from another. Along with exports, imports form the backbone of international trade. The higher the value of imports entering a country, compared to the value of exports, the more negative that country's balance of trade becomes.^[3]

PROBLEM OF STATEMENT

Foreign reserves acts as a buffer against capital outflows in excess of the trade balance. This makes foreign reserves management secondary to macroeconomic objectives, as liquidity is always the target. This also enables the monetary authority intervene in the foreign exchange market at any given time. Holding foreign reserves under both fixed and floating exchange rate regimes also acts as a “shock absorber” in terms of fluctuations in international transactions, such as variations in imports resulting from trade shocks, or in the capital account due to financial shocks.

OBJECTIVE

1.      To determine the long term relationship by utilizing Johansen Cointegration Test

2.      To explore the short term relationship by those relationship

3.      To determine the factor influence the reserve of Malaysia

2.0  LITERATURE REVIEW

Reserves are used to intervene in the foreign exchange market to influence the exchange rate. Foreign reserves are used to support monetary and foreign exchange policies, in order to meet the objectives of safeguarding currency stability and the normal functions of domestic and external payment systems (Irefin & Yaaba, 2011). Based on the literature in the (Marc-Andre & Parent, 2005), the determinants of reserve holdings reported in the literature can be grouped into five categories: economic size, current account vulnerability, capital account vulnerability, exchange rate flexibility, and opportunity cost. The table 2 below are the lists potential explanatory variables for each of these categories (Heller and Khan 1978; Edwards 1985; Lizondo and Mathieson 1987; Landell-Mills 1989; and Lane and Burke 2001).

In the literature, GDP and GDP per capita are used as indicators of economic size. The vulnerability of the current account can be captured by such measures as trade openness and export volatility. In the long run, central banks will increase their reserves in response to a greater exposure to external shocks. For this reason, the level of reserves should be positively correlated with an increase both exports and imports (Irefin & Yaaba, 2011).

According to Frenkeland Jovanovic (1981) most of the rules for a country’s demand for foreign exchange reserves consider real variables, such as imports, exports, foreign debt, severity of possible trade shocks and monetary policy considerations (Irefin & Yaaba, 2011). Similarly, Shcherbakov (2002) states that, there are some common indicators that are used to determine the adequate level of foreign reserves for an economy. According to him, some of these indicators determine the extent of external vulnerability of a country and the capability of foreign reserves to minimize this vulnerability. These indicators includes: import adequacy, debt adequacy and monetary adequacy (Irefin & Yaaba, 2011). The traditional and most prominent factor considered in determining foreign reserves adequacy is the ratio of foreign reserves to imports (import adequacy). This represents the number of months of imports for which a country could support its current level of imports, if all other inflow and outflow stops. As a rule of thumb, countries are to hold reserves in order to cover their import for three to four months (Irefin & Yaaba, 2011).

For a developing economy like Nigeria, there is need to extend the model to incorporate other variables that are peculiar in the determination of reserves holdings. Hence, variables such as Gross Domestic Product (GDP), imports, monetary policy rate which is an anchor of monetary policy and exchange rate are included in the estimation equation (Irefin & Yaaba, 2011). Thus, the equation becomes:

 Log Rt = B0 + B1logYt + B2logIMt + B3logMPRt +B4logEXRt + ut

Where R= foreign reserves, Y = Gross Domestic Product, IM = Import, MPR = Monetary Policy Rate and EXR = Exchange Rate.

The justification for including additional variables for Nigeria is that, for instance, reserves holdings are positively related with the level of international transactions hence the importance of variables such as imports and exchange rate.

3.0 METODOLOGY

DATA SOURCE AND DESCRIPTION

There are totally three variables are being used in this study, which are total reserve, export and import. I utilized 235 observations monthly in selected time series data that span from 1994 to 2013. All the data can be obtained from World Bank website.^[4]

The major variables for which data collected are defined below:

Total reserves: Total reserves comprise holdings of monetary gold, special drawing rights, reserves of IMF members held by the IMF, and holdings of foreign exchange under the control of monetary authorities. The gold component of these reserves is valued at year-end (December 31) London prices.[5] While, Foreign reserves (R) is the total assets of central bank held in different reserves currencies abroad. The reserves currencies includes; US dollar, Pound Sterling, Euro, Japanese Yen etc. The common scale variables used in the model are GDP and imports (Irefin & Yaaba, 2011).

Export: The term export means shipping the goods and services out of the port of a country. Export of commercial quantities of goods normally requires involvement of the customs authorities in both the country of export and the country of import

Import: Imports are the total monetary value of goods and services imported into the country on quarterly basis (Irefin & Yaaba, 2011).

ECONOMETRIC MODEL

R = B₁ + B₂X + B₃M + E

R = Total reserve

X = Export

M = Import

E = Error term

To check for the stationary properties of the data, the Augmented Dickey-Fuller (ADF) (Dickey and Fuller, 1979) unit root test will be applied. Next I will examine the long-run equilibrium relationship of the variables using the Johansen and Juselius (1990) cointegration test. The vector error correction estimate will then be utilized to scrutinize the impact of each of the independent variable towards the dependent variable in the model as stated in Equation (1). And I will use ARCH and GARCH family to test whether the variable is fit to with one model.

Based on literature the Autoregressive Distributed Lag (ARDL) model developed by Pesaran et al (2001) is deployed to estimate Frenkel and Jovanovic‟s “buffer stock” econometric model, but with a slight modification. The choice of ARDL is based on several considerations. First, the model yields consistent estimates of the long run normal coefficients irrespective of whether the underlying regressors are stationary at I(1) or I(0) or a mixture of both. In other words, it ignores the order of integration of the variables (Pesaran et al, 2001). Secondly, it provides unbiased estimates of the long run model as well as valid t-statistics even when some of the regressors are endogenous (Harris & Sollis, 2003). Thirdly, it has good small sample properties. In other words it yields high quality results even if the sample size is small (Irefin & Yaaba, 2011).

THEORETICAL FRAMEWORK

         

4.0 RESULT AND FINDING

ORDINARY LEAST SQUARE (OLS)

Dependent Variable: LOG(R)
Method: Least Squares
Date: 12/12/14   Time: 20:49
Sample: 1994M01 2013M10
Included observations: 238
Variable	Coefficient	Std. Error	t-Statistic	Prob.
C	11.53195	0.021664	532.3059	0.0000
LOG(X)	11.82853	0.724507	16.32633	0.0000
LOG(M)	-9.399178	0.816743	-11.50812	0.0000
R-squared	0.887677	Mean dependent var		10.84844
Adjusted R-squared	0.886722	S.D. dependent var		0.660280
S.E. of regression	0.222230	Akaike info criterion		-0.157687
Sum squared resid	11.60570	Schwarz criterion		-0.113919
Log likelihood	21.76476	Hannan-Quinn criter.		-0.140048
F-statistic	928.5946	Durbin-Watson stat		0.318120
Prob(F-statistic)	0.000000

Table 1.1

HO: X not influence R

H1: X influence R

H0: M does not influence R

H1: M does influence R

Based on the Table 1.1 above we can reject the null hypothesis with 1 percent level of significant where the prob (t-statistic) is 0.0000 is less than 0.01 which is represent highly significant relationship between X and R. Thus, we can accept the alternate hypothesis. Based on table 1.1 the prob(t-statistic = 11.50812) is more than critical value 2 and the prob= 0.00 is highly significant representing the relationship between M and R. Thus, we can reject null hypothesis and accept the alternate hypothesis.R-squared is 0.887677 meaning the 88.7677 percent of reserve is explained by export and import and the balance 11.2323 percent is explain by other factors, thus, they are not included in the model.Durbin Watson is 0.318120 which represents the data has autocorrelation and it is non-stationary thus the unit root test is needed to make the data stationary with integrating first level or second level.

AUGMENTED DICKY-FULLER (ADF) TEST STATISTIC

	R			X			M
	level	1st differ	2nd differ	Level	1st differ	2nd differ	Level	1st differ	2nd differ
Augmented Dickey-Fuller test statistic	0.204428	-9.545826		-0.576868	-6.280101		-0.629085	-6.092040
Test Critical Value:     1%	-3.457984	-3.457984		-3.458225	-2.574882		-3.458225	-2.574882
5%	-2.873596	-2.873596		-2.873701	-1.942188		-2.873701	-1.942188
10%	-2.573270	-2.573270		-2.573327	-1.615795		-2.573327	-1.615795

Table 2.1

Based on table 2.1, the R variable is stationary at I (1)because we can reject the null hypothesis (Ho) with 1 percent, 5 percent, 10 percent level of significant and the augmented dickey fuller t absolute value (9.5458260) is more than the t-values(3.457984),( 2.873596) and (2.573270) at integrated level 1 I (1). We can conclude that the data are stationary at I (1).

For Z variable, we can reject the null hypothesis (Ho) with 1 percent, 5 percent, 10 percent level of significant because the augmented dickey fuller t absolute value (6.280101) is more than the t-value, (2.574882), (1.942188) and (1.615795) at integrated level 1 I (1). We can conclude that the data is stationary at I (1). 

For M variable we can reject the null hypothesis (Ho) with 1 percent, 5 percent, 10 percent level of significant because the augmented dickey fuller t absolute value (6.092040) is more than the t-value, (2.574882), (1.942188) and (1.615795)at integrated level 1 I (1). We can conclude that the data is stationary at I (1).

Ordinary Least Square (OLS) After The Data is Stationary

Table 2.2 shows the result after we run the ordinary least squares by using integrated variable the Durbin Watson become more than critical value (2) then we conclude the model is no autocorrelation. After the data being stationary the Durbin Watson increases than before. So, we can conclude that the model has no autocorrelation problem because the Durbin Watson stat is 2.899154 which is more than critical value 2.As well the R-square is strongly indicating that the X and M are 100 percent explaining the R dependent variable.

Dependent Variable: DR
Method: Least Squares
Date: 12/12/14   Time: 21:22
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
C	-2.26E-19	1.40E-19	-1.609466	0.1089
DX	1.000000	1.55E-17	6.43E+16	0.0000
DM	-2.26E-16	1.73E-17	-13.07836	0.0000
R-squared	1.000000	Mean dependent var		0.001547
Adjusted R-squared	1.000000	S.D. dependent var		0.013810
S.E. of regression	2.14E-18	Sum squared resid		1.07E-33
F-statistic	4.89E+33	Durbin-Watson stat		2.899154
Prob(F-statistic)	0.000000

Table 2.2

The Comparison Before and After The Data Being Stationary

Graph 1, 3, and 5 is shown the data before stationary and we can see it was randomly walk. Meanwhile, the graph 2, 4, and 6 shows the variable after being stationary and converted into integrated first difference.

JOHANSEN COINTEGRATION TEST

Unrestricted Co-integration Rank Test (Trace)
Hypothesized		Trace	0.05
No. of CE(s)	Eigenvalue	Statistic	Critical Value	Prob.**
None	0.066398	19.02628	29.79707	0.4910
At most 1	0.012866	3.017952	15.49471	0.9661
At most 2	2.61E-06	0.000607	3.841466	0.9820
Trace test indicates no cointegration at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values

Table 3.1

From table 3.1 at hypothesized, number of co-integration none is show

H0: There is no co-integration equation.

H₁: There is co-integration equation.

The estimated trace statistics value of 19.02628 is less than critical value of MacKinnon at 5% significance level (29.79707 P=0.4910). This shows there is no one cointegrating equation between R, X, and M. This means there is no a long run relationship between R, X, and M.

Unrestricted Cointegration Rank Test (Maximum Eigenvalue)
Hypothesized		Max-Eigen	0.05
No. of CE(s)	Eigenvalue	Statistic	Critical Value	Prob.**
None	0.066398	16.00833	21.13162	0.2244
At most 1	0.012866	3.017345	14.26460	0.9456
At most 2	2.61E-06	0.000607	3.841466	0.9820
Max-eigenvalue test indicates no co-integration at the 0.05 level
* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values

Table 3.2

From table3.2, the max-Eigen statistics also shows the number of co-integration equation. The hypothesized number of co-integration equation for  of none shows there is no co-integration equation. At this level, the estimated max-Eigen statistics of 16.00833 is smaller than the MacKinnon critical value of 21.13162 at 5% significance level (p=0.2244). This means there is no a long run relationship between R, X, and M.

LONG RUN MODEL

Normalized co-integrating coefficients (standard error in parentheses)
R	X	M
1.000000	-960761.5	771035.5
	(141723.)	(156231.)

R = – 960761.5 + 771035.5

(141723.)           (156231.)

Estimated-t

X960761.5/141723 = 6.7791

M771035.5/156231 = 6.8353

The estimated t-value for X (6.7791) is greater than critical t-value of t (2). Therefore x is significant in explaining the changes in y at 5% significance level.

The estimated t-value of M (6.8353) is greater than critical t-value of t (2). Thus M can explain the changes in y at 5% significance level. Therefore, X, and M can influence R in long run.

Unrestricted Var

Vector Auto-regression Estimates
Date: 12/13/14   Time: 08:27
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Standard errors in ( ) & t-statistics in [ ]
	R	X	M
R(-1)	1.399820	9.64E-07	9.27E-07
	(0.05975)	(3.7E-07)	(3.4E-07)
	[ 23.4298]	[ 2.60085]	[ 2.75684]
R(-2)	-0.379027	-4.97E-07	-6.32E-07
	(0.06324)	(3.9E-07)	(3.6E-07)
	[-5.99361]	[-1.26555]	[-1.77562]
X(-1)	18022.96	0.643799	0.224729
	(16499.7)	(0.10237)	(0.09290)
	[ 1.09232]	[ 6.28907]	[ 2.41910]
X(-2)	-43659.39	-0.058098	-0.440797
	(15929.0)	(0.09883)	(0.08968)
	[-2.74087]	[-0.58787]	[-4.91495]
M(-1)	-2973.498	0.364043	0.741475
	(17243.6)	(0.10698)	(0.09709)
	[-0.17244]	[ 3.40279]	[ 7.63728]
M(-2)	25502.70	-0.039125	0.411956
	(17675.5)	(0.10966)	(0.09952)
	[ 1.44283]	[-0.35677]	[ 4.13950]
C	1072.199	0.040762	0.032412
	(1933.61)	(0.01200)	(0.01089)
	[ 0.55451]	[ 3.39778]	[ 2.97718]
R-squared	0.997519	0.993035	0.993022
Adj. R-squared	0.997454	0.992853	0.992840
Sum sq. resids	9.53E+08	0.036680	0.030207
S.E. equation	2039.886	0.012656	0.011485
F-statistic	15343.78	5441.673	5431.706
Log likelihood	-2129.790	699.9152	722.8249
Akaike AIC	18.10839	-5.872163	-6.066312
Schwarz SC	18.21113	-5.769422	-5.963572
Mean dependent	63846.25	0.858027	0.880401
S.D. dependent	40425.30	0.149700	0.135727
Determinant resid covariance (dof adj.)		0.031293
Determinant resid covariance		0.028590
Log likelihood		-585.1541
Akaike information criterion		5.136899
Schwarz criterion		5.445122

Table 4.1

We have to run unrestricted Var because there is no relationship in long run for this model. Base on table 4.1, can conclude the R is has short term relationship with X and M because the t-stat 2.60085 and 2.75684 is more than critical value (2).

Residual Test

Graph 7

Base on graph 7 above the residual is low volatile for some period of time and continued with also low volatile for another period of time then we can justify that an ARCH and GARCH family model can be used.

Arch Effect Test

Heteroskedasticity Test: ARCH
F-statistic	0.319847	Prob. F(1,233)		0.5722
Obs*R-squared	0.322150	Prob. Chi-Square(1)		0.5703
Test Equation:
Dependent Variable: RESID^2
Method: Least Squares
Date: 12/14/14   Time: 00:42
Sample (adjusted): 1994M04 2013M10
Included observations: 235 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
C	9.87E-36	3.59E-36	2.747913	0.0065
RESID^2(-1)	0.037024	0.065465	0.565550	0.5722
R-squared	0.001371	Mean dependent var		1.02E-35
Adjusted R-squared	-0.002915	S.D. dependent var		5.40E-35
S.E. of regression	5.41E-35	Sum squared resid		6.82E-67
F-statistic	0.319847	Durbin-Watson stat		2.000600
Prob(F-statistic)	0.572244

Table 5.1

H₀: No ARCH effect

H₁: There is Arch effect

We cannot reject the hypothesis null because the prob chi-square is 0.572244 is more than 5 percent level of significant than the data is fail to reject that there is no ARCH effect   in our model.

SERIAL CORRELATION TEST

Date: 12/14/14   Time: 01:02
Sample: 1994M03 2013M10
Included observations: 236

Arch Effect Test

Heteroskedasticity Test: ARCH
F-statistic	0.059602	Prob. F(1,233)		0.8073
Obs*R-squared	0.060099	Prob. Chi-Square(1)		0.8063
Test Equation:
Dependent Variable: WGT_RESID^2
Method: Least Squares
Date: 12/14/14   Time: 01:16
Sample (adjusted): 1994M04 2013M10
Included observations: 235 after adjustments
Variable	Coefficient	Std. Error	t-Statistic	Prob.
C	1.028378	0.348523	2.950680	0.0035
WGT_RESID^2(-1)	-0.015991	0.065501	-0.244136	0.8073
R-squared	0.000256	Mean dependent var		1.012216
Adjusted R-squared	-0.004035	S.D. dependent var		5.234924
S.E. of regression	5.245475	Akaike info criterion		6.161083
Sum squared resid	6410.997	Schwarz criterion		6.190526
Log likelihood	-721.9272	Hannan-Quinn criter.		6.172953
F-statistic	0.059602	Durbin-Watson stat		2.000226
Prob(F-statistic)	0.807340

Table 5.3

H0: There is no ARCH effect

H1: There is ARCH effect

We have to accept the null hypothesis with 5 percent level of significant because the prob (0.8073) is more than 5 percent level of significant.

NORMALLY DISTRIBUTED TEST

Graph 8

H₀: The data is not normally distributed

H₁: The data is normally distributed

The Jarque-Bera prob is less than 5 percent of level of significant, thus, we can reject the null hypothesis and concluded the data is normally distributed.

The Fitted Model Test D(R)

GARCH

Dependent Variable: DR
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/15/14   Time: 02:27
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 14 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(1) + C(2)*RESID(-1)^2 + C(3)*GARCH(-1)
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C	1.98E-05	4.77E-06	4.144316	0.0000
RESID(-1)^2	0.072293	0.022838	3.165519	0.0015
GARCH(-1)	0.835157	0.044685	18.68974	0.0000
R-squared	-0.012599	Mean dependent var		0.001547
Adjusted R-squared	-0.021291	S.D. dependent var		0.013810
S.E. of regression	0.013956	Akaike info criterion		-5.753751
Sum squared resid	0.045380	Schwarz criterion		-5.709719
Log likelihood	681.9426	Hannan-Quinn criter.		-5.736001
Durbin-Watson stat	1.816389

Table 5.4

EGARCH

Dependent Variable: DR
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/15/15   Time: 02:28
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 35 iterations
Presample variance: backcast (parameter = 0.7)
LOG(GARCH) = C(1) + C(2)*ABS(RESID(-1)/@SQRT(GARCH(-1))) + C(3)
*LOG(GARCH(-1))
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C(1)	-1.378363	0.281393	-4.898363	0.0000
C(2)	0.167184	0.058467	2.859455	0.0042
C(3)	0.850599	0.028905	29.42758	0.0000
R-squared	-0.012599	Mean dependent var		0.001547
Adjusted R-squared	-0.021291	S.D. dependent var		0.013810
S.E. of regression	0.013956	Akaike info criterion		-5.781030
Sum squared resid	0.045380	Schwarz criterion		-5.736999
Log likelihood	685.1616	Hannan-Quinn criter.		-5.763281
Durbin-Watson stat	1.816389

Table 5.5

TARCH

Dependent Variable: DR
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/14/13   Time: 02:30
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 83 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(1) + C(2)*RESID(-1)^2 + C(3)*RESID(-1)^2*(RESID(-1)<0) +
C(4)*GARCH(-1)
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C	5.65E-05	1.02E-05	5.518783	0.0000
RESID(-1)^2	0.139978	0.048400	2.892081	0.0038
RESID(-1)^2*(RESID(-1)<0)	0.860897	0.437444	1.968016	0.0491
GARCH(-1)	0.439916	0.096858	4.541889	0.0000
R-squared	-0.012599	Mean dependent var		0.001547
Adjusted R-squared	-0.025693	S.D. dependent var		0.013810
S.E. of regression	0.013986	Akaike info criterion		-5.785566
Sum squared resid	0.045380	Schwarz criterion		-5.726857
Log likelihood	686.6968	Hannan-Quinn criter.		-5.761900
Durbin-Watson stat	1.816389

Table 5.6

GARCH

Akaike info criterion               -5.753751

Schwarz criterion                    -5.709719

EGARCH

Akaike info criterion               -5.781030

Schwarz criterion                    -5.736999

TARCH

Akaike info criterion               -5.785566

Schwarz criterion                    -5.726857

We can conclude that GARCH is the best model for D(R) because it has the lower AIC and SIC.

The Fitted Model Test D(X)

GARCH                                                                            

Dependent Variable: DX
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/14/13   Time: 02:17
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 14 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(1) + C(2)*RESID(-1)^2 + C(3)*GARCH(-1)
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C	1.98E-05	4.77E-06	4.144316	0.0000
RESID(-1)^2	0.072293	0.022838	3.165519	0.0015
GARCH(-1)	0.835157	0.044685	18.68974	0.0000
R-squared	-0.012599	Mean dependent var		0.001547
Adjusted R-squared	-0.021291	S.D. dependent var		0.013810
S.E. of regression	0.013956	Akaike info criterion		-5.753751
Sum squared resid	0.045380	Schwarz criterion		-5.709719
Log likelihood	681.9426	Hannan-Quinn criter.		-5.736001
Durbin-Watson stat	1.816389

Table 5.7

EGARCH

Dependent Variable: DX
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/14/13   Time: 02:19
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 35 iterations
Presample variance: backcast (parameter = 0.7)
LOG(GARCH) = C(1) + C(2)*ABS(RESID(-1)/@SQRT(GARCH(-1))) + C(3)
*LOG(GARCH(-1))
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C(1)	-1.378363	0.281393	-4.898363	0.0000
C(2)	0.167184	0.058467	2.859455	0.0042
C(3)	0.850599	0.028905	29.42758	0.0000
R-squared	-0.012599	Mean dependent var		0.001547
Adjusted R-squared	-0.021291	S.D. dependent var		0.013810
S.E. of regression	0.013956	Akaike info criterion		-5.781030
Sum squared resid	0.045380	Schwarz criterion		-5.736999
Log likelihood	685.1616	Hannan-Quinn criter.		-5.763281
Durbin-Watson stat	1.816389

Table 5.8

TARCH

Dependent Variable: DX
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/14/13   Time: 02:20
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 83 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(1) + C(2)*RESID(-1)^2 + C(3)*RESID(-1)^2*(RESID(-1)<0) +
C(4)*GARCH(-1)
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C	5.65E-05	1.02E-05	5.518783	0.0000
RESID(-1)^2	0.139978	0.048400	2.892081	0.0038
RESID(-1)^2*(RESID(-1)<0)	0.860897	0.437444	1.968016	0.0491
GARCH(-1)	0.439916	0.096858	4.541889	0.0000
R-squared	-0.012599	Mean dependent var		0.001547
Adjusted R-squared	-0.025693	S.D. dependent var		0.013810
S.E. of regression	0.013986	Akaike info criterion		-5.785566
Sum squared resid	0.045380	Schwarz criterion		-5.726857
Log likelihood	686.6968	Hannan-Quinn criter.		-5.761900
Durbin-Watson stat	1.816389

Table 5.9

GARCH

Akaike info criterion               -5.753751

Schwarz criterion                    -5.709719

EGARCH

Akaike info criterion               -5.781030

Schwarz criterion                    -5.736999

TARCH

Akaike info criterion               -5.785566

Schwarz criterion                    -5.726857

We can concluded that GARCH is the best model for D(X) because it has the lower AIC and SIC.

THE FITTED MODEL TESTD (M)

GARCH

Dependent Variable: DM
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/14/13   Time: 02:38
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 23 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(1) + C(2)*RESID(-1)^2 + C(3)*GARCH(-1)
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C	0.000226	6.79E-06	33.36290	0.0000
RESID(-1)^2	0.001386	0.001321	1.049360	0.2940
GARCH(-1)	-1.000169	0.001424	-702.5541	0.0000
R-squared	-0.010949	Mean dependent var		0.001297
Adjusted R-squared	-0.019626	S.D. dependent var		0.012423
S.E. of regression	0.012544	Akaike info criterion		-6.189925
Sum squared resid	0.036665	Schwarz criterion		-6.145893
Log likelihood	733.4112	Hannan-Quinn criter.		-6.172176
Durbin-Watson stat	1.943615

Table 6.0

EGARCH

Dependent Variable: DM
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/14/13   Time: 02:39
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 21 iterations
Presample variance: backcast (parameter = 0.7)
LOG(GARCH) = C(1) + C(2)*ABS(RESID(-1)/@SQRT(GARCH(-1))) + C(3)
*LOG(GARCH(-1))
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C(1)	-4.290583	1.105587	-3.880818	0.0001
C(2)	0.279081	0.095713	2.915822	0.0035
C(3)	0.529858	0.122047	4.341415	0.0000
R-squared	-0.010949	Mean dependent var		0.001297
Adjusted R-squared	-0.019626	S.D. dependent var		0.012423
S.E. of regression	0.012544	Akaike info criterion		-5.936776
Sum squared resid	0.036665	Schwarz criterion		-5.892745
Log likelihood	703.5396	Hannan-Quinn criter.		-5.919027
Durbin-Watson stat	1.943615

Table 6.1

TARCH

Dependent Variable: DM
Method: ML - ARCH (Marquardt) - Normal distribution
Date: 12/14/13   Time: 02:40
Sample (adjusted): 1994M03 2013M10
Included observations: 236 after adjustments
Convergence achieved after 8 iterations
Presample variance: backcast (parameter = 0.7)
GARCH = C(1) + C(2)*RESID(-1)^2 + C(3)*RESID(-1)^2*(RESID(-1)<0) +
C(4)*GARCH(-1)
Variable	Coefficient	Std. Error	z-Statistic	Prob.
	Variance Equation
C	0.000151	1.95E-05	7.749871	0.0000
RESID(-1)^2	0.117887	0.078832	1.495422	0.1348
RESID(-1)^2*(RESID(-1)<0)	-0.133214	0.102083	-1.304959	0.1919
GARCH(-1)	-0.000121	0.124311	-0.000975	0.9992
R-squared	-0.010949	Mean dependent var		0.001297
Adjusted R-squared	-0.024021	S.D. dependent var		0.012423
S.E. of regression	0.012571	Akaike info criterion		-5.916491
Sum squared resid	0.036665	Schwarz criterion		-5.857782
Log likelihood	702.1459	Hannan-Quinn criter.		-5.892825
Durbin-Watson stat	1.943615

Table 6.2

GARCH

Akaike info criterion               -6.189925

Schwarz criterion                    -6.145893

EGARCH

Akaike info criterion               -5.936776

Schwarz criterion                    -5.892745

TARCH

Akaike info criterion               -5.916491

Schwarz criterion                    -5.857782

We can concluded that TARCH is the best model for D(M) because it has the lower AIC and SIC.

5.0 CONCLUSION

We found that the data is not stationary in the beginning and we have to integrate it into first difference to make it stationary for our analysis. Then after the data become stationary the export and import highly significantly explaining the total reserve Malaysia then we can suggest to the policy maker that in order to increase the total revenue we have to increase the export. Reducing imports is important as well because it negatively influence the total reserve Malaysia. But in the long run we found that import and export do not have relationship with Malaysia total reserve but in short run they will influence the total reserve and other than that we can choose the GARCH model for reserve and export variable and meanwhile TARCH model for import variable because the data is fitted with those model. And lastly for future researcher these paper is provide a latest analysis with the newest data and updated methodology and it will contributed to the body of knowledge.