Practical part of this thesis will analyze the relationship FDI net inflows and real GDP growth as well as the relationship between DS and FDI of Kazakhstan. Data used in this paper are yearly figures covering the 1993-2019 period. The data base is the World Bank (2019), National Bank of Republic of Kazakhstan (2019) and Agency of Statistics of Kazakhstan (2019). FDI and DS are expressed in absolute values. GDP growth is express in both absolute and percentage (natural logarithmic form) forms. All results shown in this work were assessed with statistical data analysis and table-chart overview.
2.1.Methodology
Methodology of practical part will partially include analyses the model of “FDI, DS and Economic Growth in Kazakhstan: co-integration and causality model” run by A. Naraliyeva and S.T. Katirciouglu, 2006. As a main tests ADF17 and PP18 unit root tests were employed.
Also, “FDI and Economic growth in Malaysia” by M.S. Karimi and Z.Yusop (2009) who tested same connection but with the sample of Malaysia.
17 The Augmented Dickey-Fuller, tests the null hypothesis that a unit root is present in a time series sample.
18 Philips-Perron used in time series analysis to test the null hypothesis that a time series is integrated of order.
The Augmented Dickey-Fuller (ADF) and Root Tests of the Phillips-Perron unit are used to test the degree of integration and potential co-integration between the variables (Dickey and Fuller, 1981; Phillips and Perron, 1988). To test the existence of unit root process in the series the general model was used,
∆𝑦!= 𝑎"+ 𝛾𝑦!#$+ 𝑎%𝑡 + ∑(')%𝛽&∆𝑦!#'#$+∈!, (1)
Where 𝑦- is time series, t- time (trend factor), 𝑎- constant term, ∈!- Gaussian white noise19, p-the lag order. The Akaike Information Criterion (AIC) has chosen the number of 'p' lags in the dependent variable to ensure that the errors are white noise. One problem with the inclusion of the extra calculated parameters is that the degrees of freedom and test power are decreased.
The null hypothesis of a unit root (𝛾=0) might fail to reject due to misspecification regarding the following part of the regression. Consequently, when the type of the data generation method is uncertain, Doldado, Jenkinson and Sosvilla-Rivero (1990) also suggest starting from the most general model to test for a unit root. The basic principle is to select a configuration which, under both null and alternative hypotheses, is a plausible explanation of the data. The power of the test will go to zero if the intercept or time pattern is wrongly omitted. "Reduced power means the researcher will conclude that when, in fact, none is present, the process contains a unit root."20 If the variables are stationary, a linear combination of integrated variables is said to be co-integrated. Such co-integrating partnerships require multiple economic models.
Kwiatkowski Phillips, Schmidt and Shin's test (1992) (KPSS) is proposed to remove a potential low power against stationary close unit root processes that occur in the ADF and PP tests to validate the test results obtained from the ADF and PP tests. The KPSS test complements the ADF and PP tests, where the KPSS test's null hypothesis is that a sequence is stationary. The KPSS test is built on the presumption that, in the following equation, a sequence can be investigated in three parts: a time trend, a random walk, and a stationary error:
19 Any two values of GWN are statistically independent now matter how close they are in time.
20 Enders, 1995: p. 255.
𝑦!= 𝜌𝑡 + 𝑟𝑤!+ 𝜀!, (2)
Where 𝑟𝑤!= 𝑟𝑤!#$+ 𝑣! and 𝑣! is i.i.d. (0, 𝛿*%). The above regression can essentially be performed in two ways: first with a constant in the case of the stationary phase, second with both a constant and a trend in the case of the stationary trend. In the following equation, we then use the 𝜀! residuals from the regression and calculate the LM statistics:
As in the following equation, 𝑉!% can be constructed to be a more robust estimator because of the implications of 𝜀! ‘s behavior, as in Phillips and Perron's (1988) paper:
𝑉!(𝑝) = 𝑇"#' 𝜀$!
Where w(v,p) is a variable function. According to Newey (1987) it is possible to use the Bartlett window as w(v,p)=1-v/(v+1). After this observation, the test statistics of KPSS test is:
𝑡 = 𝑇#%5 𝑉!%
,
')$
∕ 𝑉%(𝑝)
Once the integration order is determined and if the array is in the same integration order, it is then important to evaluate co-integration between the variables to identify any long-term (3)
(4)
(5)
(6)
relationship. For co-integration, the trace test is more reliable than the overall private value test. The trace test aims to calculate the number of vectors co-integrating between variables.
For potential co-integration, there should be at least one co-integrating vector. This method makes it possible to estimate all possible co-integrating vectors between the set of variables and is the most precise test to prevent the issues stemming from the procedure of Engel and Granger (1987)21. After this consideration VAR model is:
𝑋! = Π$𝑋!#$+. . . +Π-𝑋!#- + 𝜇 + 𝑒! (for t = 1, ...T), (7)
Where 𝑋!, 𝑋!#$, …, 𝑋!#- are the current and lagging value vectors of P variables in the model. Π$, Π- are coefficient matrix with dimensions. μ is an intercept vector which shows dummy variables too, which guarantees that errors e are white noise. 𝑒! is a random error vector. In practice, the number of lagged values is calculated in a way that the error terms are not auto-correlated significantly. If we add Π$ 𝑋!#%, … , Π-#$X!#- and 𝑋!#$, …, 𝑋!#- to both sides of the VAR model, it will be represented as:
∆𝑋!=Γ$∆𝑋!#$ +…+ Γ-#$∆𝑋!#-/$ + Π𝑋!#-+𝜇+𝑒!, (8) Where 𝛤'=-(I-𝛱'-…-𝛱'); (i =1,2, …, K-1); 𝛱=-(I-𝛱$-…- 𝛱$) where I is denoted as identity matrix. The level of the coefficient matrix 𝛱 provides the number of long-run correlations between the system's variables. During the test three possible cases22 were stated:
1. If the ranks are equal to P[r(𝛱) =P] indicating that 𝛱 has maximum rank, then every linear I(1) sequence combination is stationary.
2. If the rank is empty or equal to zero (r(𝛱) =0, i.e. 𝛱 is a 0 matrix), then there is no relationship of co-integration.
3. If the rank is between P(0 < r(𝛱)< P) and 0, then there’re a and b matrices with (pxr) dimension, that 𝛱=ab can be defined. Matrix β is referred to as the 'co-integrating matrix' while matrix a is referred to as the 'matrix of adaptation' or the 'matrix of input.'
21 For remarks on the drawbacks of the Engel and Granger (1987) approach compared with the cointegration technique of Johansen and Juselius (1990)
22 Johannes and Juselius (1990)
The rank of 𝛱 is the sum of co-integrated relationships (i.e. r) calculated by checking whether its own values (λ') differ from zero statistically. Observations obtained indicate that the use of the eigen values of 𝛱 listed from the largest to the smallest is intended for trace calculation.
Following formula determines the trace statistics (λ!0123):
λ!0123 = −𝑇 ∑ 𝐿𝑛(1 − 𝜆'), i= r+1, …, n-1 (9)
With given hypothesis:
𝐻": r=0, 𝐻$: r ³ 1 𝐻": r£1, 𝐻$: r ³ 2 𝐻": r£2, 𝐻$: r ³ 3
The discovery that a unit root can contain several macro time series has stimulated the development of the non-stationary time series analysis theory. Empirical studies have indicated that in the time series examined, the presence of non-stationarity can lead to misleading results of regression and disprove the results obtained using Granger Causality.
A long-term causal association could be inferred between non-stationary time series when they are co-integrated, but if co-integration analysis is excluded, causality tests present proof of simultaneous associations rather than causal relationships between variables.
When co-integrating vectors are obtained in the series, the simple Granger's causality test becomes insufficient, since the results of co-integration suggest that the series has the following interpretations of error correction. These are important for the simple Granger causality test to be supplemented with the ECM23, derived from the residuals of the sufficient co-integration relationship to the causality test:
∆ ln 𝑌! = 𝐶"+ 5 𝛽'∆ ln 𝑌!#'
4
')$
+ 5 𝛼'∆ ln 𝑋!#'+ 𝑝'
4
')$
𝐸𝐶𝑇!#$+ 𝑢! ,
23Error Correction Mechanism
(10)
∆ ln 𝑋! = 𝐶"+ 5 𝛾'∆ ln 𝑋!#'
Where Y and X variables are under consideration, 𝑝' is coefficient of adjustment, 𝐸𝐶𝑇!#$ is error correction term for the equation of growth, ∆ is first difference operator indicator. In equation (10), Granger X causes Y if 𝛼' and 𝑝' vary significantly from 0. In equation (11),
Where 𝜇! and 𝑣! are serially white-noise residuals that are uncorrelated, p, q, s and r are lag length for every variable in all equations.
2.2. Data
The main aim of this section is to show how the results on FDI, and Economic Growth were obtained. Analysis of the real GDP growth, DS and FDI was run in a yearly basis to outline the results of yearly lags, whereas regional FDI comparison and completion of Population, GDP per capita growth and FDI relationship were tested with the quarterly data provided by Agency of Statistics of Kazakhstan and the World Bank. The first FDI inflow was contributed to the economy of Kazakhstan was made in 1991, however earliest data on DS was available from 1993, so the tested period that was chosen in this model is from 1993 to 2019. Also, as the general model for Kazakhstan contains more variables and items compared to the regional model, it is better to use yearly data to avoid the influence seasonal and cyclical factors on estimated values of the regression models.
(11)
(12)
(13)
Variables selected for the analysis are:
- Foreign Direct Investment (FDI) - GDP growth rate
- Real GDP (Gross Domestic Product) - Population
- FDI per capita - GDP per capita
- Domestic Savings (DS)
All variables except real GDP growth and Population are represented in millions of US dollars. Real GDP growth is measured in billions of US dollars and annual % points. FDI is shown as a share of total GDP to test the instability effect on economic growth. Regression maps made for regional comparison of FDI shares are done for the period of 2014-2019 with the main variable- net FDI inflow. Same data applied to calculate the Moran’s index. GDP per capita for each region was calculated by the 12.2019 exchange rate where 1USD
=378,97KZT, all data was transferred from KZt to US$.
For the computation of 1,2- and 3-year lag of GDP, annual GDP growth rate (%) was transferred into 3-year smoother averages, further, to be shown in a trendline. Same process was applied to the FDI as % of GDP growth rate variable. Both variables were transferred from absolute numbers to relative percentage points for better vision of plot area. Period chosen for this analysis was from 1992 to 2019.
Residual analysis was plotted in the maps for better view on regional development of selected areas. Statistical regression was run to test correlation and collinearity of selected variables.
For this regression dataset of net FDI per capita and GDP per capita of different regions from the period 2014-2019 was used. It examined 16 different regions, also considering oil & gas dummy, to see whether the regions that perform in fuel industry have a greater share of FDI.
2.3. Analysis
Kazakhstan’s net FDI inflows have been increasing over the years, despite the sharp declines in the amount of foreign capital inflow, by the 2019 in increased 4 times than it used to be in 2017. Following table contains data on FDI inflow in billions of USD from 1992-2019.
Figure 2: Scatterplot of net FDI inflow in Kazakhstan
Source: Own computation in Excel
FDI instability was a great topic for debate around economists of Kazakhstan, as it clearly had an impact on country’s export level and technological advancement. Both major declines in FDI happened due to financial crisis in 2010 and 2018, and government doing its best to improve regulations in favor of foreign investors.
Figure 3 shows growth of two variables: real GDP growth and net FDI inflow over the period of 1991 to 2019. Despite GDP growth decline in 1990s, FDI had stable and great amount of
% share of overall GDP growth at that times. With the gradual increase in other components of GDP it becomes harder to see FDI’s influence over years, however it was present in the years of 2006, 2008 and 2016 where after greater contribution of foreign investors into Kazakhstani property we see a small increase in GDP.
0 2 4 6 8 10 12 14 16 18 20
1990 1995 2000 2005 2010 2015 2020 2025
FDI inflow (billions of USD)
Figure 3: real GDP growth and net FDI inflow in Kazakhstan
Source: Own computation in Excel
In the next chart the relationship between the GDP growth (%) and FDI as a share of GDP (%) is plotted.
Figure 4: GDP growth and FDI as share of GDP growth
Source: own computations in Excel
If we look at year 2001 where both indicators reach their highest points for the first decade, that was the year when Atyrau region boosted in number of oil fields attracting first large
0 50 100 150 200 250
1991 1995 1999 2003 2007 2011 2015 2019
billion USD
Real GDP and FDI in Kazakhstan
FDI inf (bil USD) GDP (bil USD)
-15 -10 -5 0 5 10 15
1991 1994 1997 2000 2003 2006 2009 2012 2015 2018
GDP growth rate % FDI as % of GDP
amount of FDI. In 2009 due to post financial crisis situation GDP growth rate almost hit negative mark, when FDI same year increased up to 13% of total GDP of the country same year. Consequently, GDP growth rate of Kazakhstan increased almost 6 times compared to the previous year. Same process applies to the FDI boom in 2016, after which economic growth of the country stabilizes again, doubling in numbers. However, periodically declines in FDI (see 2010, 2017 observations) doesn’t affect GDP growth rate as sharply as they were in the beginning of economic formation, which means that currently Kazakhstan does not depend on FDI as it used to.
Figure 5: Domestic Savings and FDI inflow in Kazakhstan
Source: own computations in Excel
It is visible that FDI affect DS with one year difference, when increase happened in total FDI amount in 2015, DS also gradually increase from 2016. In 2004 FDI inflow slightly decreased, but it showed no effect in DS change. However, change in FDI and DS is not predictably related. Co-fluctuations happening in both variables can give a clear image on relationship between FDI and DS. Until 2016, any fluctuation happening in FDI affected Domestic Savings significantly, however, decrease in foreign capital inflow in 2016 didn’t lead to relatable decrease in DS amount, conversely it increased for two years in a row, which shows negative correlation between those two figures in the given time.
0
1991 1995 1999 2003 2007 2011 2015
DS (bil USD) - FDI inf (bil USD)
-Figure 6: FDI inflow in transitionary economies in billion USD
Source: Own computation in Excel
Figure 7: GDP growth in transitionary economies in billion USD
Source: Own computations in Excel
Figures 6 and 7 are presented to show the relationship between FDI and GDP growth in other transitional economies such as Armenia, Kyrgyzstan, Tajikistan, Uzbekistan, Ukraine, and Turkmenistan for the period from 1997 to 2019. Apart from Kazakhstan, Ukraine,
-2 0 2 4 6 8 10 12 14 16 18 20
1997 2001 2005 2009 2013 2017
Kazakhstan Armenia Kyrgyzstan Tajikistan Uzbekistan Ukraine Turkmenistan
0 50 100 150 200 250
1997 2001 2005 2009 2013 2017
Kazakhstan Armenia Kyrgyzstan Tajikistan Uzbekistan Ukraine Turkmenistan
Turkmenistan, and Uzbekistan have higher shares of FDI inflow among CIS transitional countries. If we compare the results obtained from Figure 6. To Figure 7, same countries are to be in top positions in terms of GDP growth, which proves that FDI influences GDP growth, despite the country specifics in terms of investment. Period between 2009 and 2017 expresses the largest ratio of FDI-GDP growth relationship between above mentioned countries.
Regions of Kazakhstan are also a large topic for the discussion in the empirical part of this work. Assumption that is going to be analyzed states that the greater is the region’s FDI share, the higher are living standards of residents, in other words the higher is GDP per capita.
Figure 8: FDI inflow in different regions of Kazakhstan from 2014 to 2019, millions of USD
Source: Own computations in Excel
As mentioned in the theoretical section, Atyrau, Almaty, East Kazakhstan, West Kazakhstan and Aktobe gather largest share of FDI inflows coming to the country. All the regions (except Almaty) specify on oil & gas production, which is the major investment sector in Kazakhstan.
If FDI per capita is high in those regions, so must be the GDP per capita. For the better vision of this relationship regression maps were used.