Professional Documents
Culture Documents
net/publication/329315962
CITATIONS READS
149 2,730
1 author:
Thabani Nyoni
University of Zimbabwe
357 PUBLICATIONS 853 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
THE IMPACT OF MACROECONOMIC VARIABLES ON STOCK MARKET DEVELOPMENT IN ZIMBABWE (1990-2018) View project
All content following this page was uploaded by Thabani Nyoni on 30 November 2018.
Thabani Nyoni
Department of Economics
University of Zimbabwe, Harare, Zimbabwe
Abstract
Using annual time series data, we study inflation in Kenya over the
period 1960 – 2017 using both ARIMA and GARCH approaches. Based on
both minimum AIC and Theil’s U, the study presents the ARIMA (2, 2, 1)
model, the ARIMA (1, 2, 0) model and the AR (1) – GARCH (1, 1) model.
Our diagnostic tests indicate that the CPI series is I (2), the residuals of
the ARIMA (2, 2, 1) and ARIMA (1, 2, 0) models are stationary and that
both the ARIMA (2, 2, 1) and the ARIMA (1, 2, 0) models are stable. The
results of this study indicate that annual inflation in Kenya is likely to
continue rising. The policy implications emanating from this study are 3 –
fold and are envisioned to assist policy makers in restoring and
maintaining price stability in Kenya.
0.0 Introduction
1
Kenya Vision 2030 is a long – term socio – economic blueprint whose main
objective is to transform Kenya into a newly industrializing, middle – income
country. If this main objective is to be achieved, then we cannot rule out the
need for policy makers to be well equipped with reliable inflation forecasts.Policy
makers, as noted by Nyoni (2018k); can get prior indication about possible
future inflation through inflation forecasting.
DIMORIANREVIEWwww.dimorianreview.comA PEER REVIEWED MULTIDISCIPLINARY INTERNATIONAL INDEXED E-JOURNAL
16
Vol-5, Issue-6 November-December 2018 eISSN: 2394-9163
A few studies have modeled and forecasted inflation in Kenya, for example;
Uwilingiyimana et al (2015), Lidiema (2017) and Fwaga et al (2017). These
studies have either employed ARIMAor GARCH approaches, separately; no study
has attempted to rely on more than one modeling technique. This paper, whose
main objective is to model and forecast inflation in Kenya; is quite different from
these existing studies in the sense that I rely on both ARIMA and GARCH
analysis in trying to model and forecast inflation in Kenya. The results of the
study are envisioned to assist policy makers in maintaining general
macroeconomic stability in Kenya which will in turn play a pivotal role in the
achievement of the much touted Kenya’s Vision 2030. The rest of the paper is
organized as follows: literature review, materials & methods, results and
conclusion & policy prescriptions; in chronological order.
The major goals of economic policies include high unemployment; and stable
and increased growth (Mutwiri, 2017). Although there is no common concession
that all these goals are compatible, there is agreement on the roles various
instruments can and should play in aiding the realization of these goals
(Friedman, 1968).It is almost needless to mention that in most economies (both
developing and developed),bothmonetary and fiscal policies compliment each
other in trying to achieve the desired economic goals. While various theories of
inflation, basically do not agree on their views on inflation, it is imperative to
note that both fiscal and monetary policies play a pivotal role in the economy.
Below is a brief overview ofthe twomost celebrated and widely used theories of
inflation nowadays:
Having been developed by the great economist, John Maynard Keynes; the
Keynesian theory argues that increase in savings will not lead to lower interest
rates, as long as the economy suffers under – employment. The Keynesian
theory avers that an increase in general price level (inflation) is caused by an
increase in aggregate demand (well, above aggregate supply). Keynesians
believe that if the economy is at full employment output level, an increase in
government expenditure, a rise in private consumption and a rise in private
investment would result in a rise aggregate demand; thereby causing inflation.
Keynes himself suggested that governments should play a more active role in
the economy – he advocated for government intervention in order to stabilize
the economy.
Employing the simplest form of the GARCH model, the GARCH (1, 1) model;
Banerjee (2017) analyzed inflation in 41 countries over the period January 1958
to February 2016 and found out that in the long – run, conditional volatility of
inflation is 3.5 times greater in developing countries compared to advanced
countries. Relying on a simple GARCH (1, 1) model, Sek & Har (2012) analyzed
inflation dynamics in three Asian countries, namely Korea, Philippines and
Thailand over the periods, September 1985 to June 2010, January 1985 to June
2010 and January 1987 to June 2010 respectively; and found out that there is
lower volatility of inflation and also persistence of inflation declines in the
countries under study.In Croatia, Zivko & Bosnjak (2017), analyzed inflation
using monthly inflation data from January 1997 to November 2015; employing
ARIMA and ARCH family models and they found out that the ARCH (1) model is
the best model for explaining CPI volatility in Croatia while the ARIMA (0, 1,
1)(0, 1, 1)12 is reliable in forecasting CPI in Croatia.In Pakistan, Saleem (2008);
looked at inflation dynamics using data over the period January 1990 to May
2007, and employed VAR, ARCH, GARCH and EGARCH methodologies and found
out that inflation is indeed volatile in nature and the time effect model is
significant.
that the behaviour of inflation over the study period follows an AR (1) – IGARCH
(1, 1) process.
From the empirical literature review, it is quite clear that most authors either use
ARIMA or ARCH family type models to model and forecast inflation. While my
basic methodological approach may be very similar to the one used by
Uwilingiyimana et al (2015), it is quite unique and still different in the sense that
I employ both ARIMA and GARCH approaches separately and then compare both
their parsimony (using the AIC values) and forecast accuracy (using the Theil’s
U).
2 Objective
Through this research article, an attempt is made to study the rend of inflation
in Kenya over the period 1960 – 2017.
3 Methodology
3.1 Model building & estimation procedures
Given:
where μt is a purely random process with mean zero and varience σ2. We say
that equation [1] is a Moving Average (MA) process of order q, commonly
denoted as MA (q). CPI is the Consumer Price Index in Kenya at time t, ɑ0 … ɑq
are estimation parameters, μt is the current error term while μt-1 … μt-q are
previous error terms. Hence:
= + ……………………………………………………….………..……………………………. [2]
and:
= + ……………………...…………………………………………..……………………….. [8]
or as:
where:
Given:
or as:
where:
2
Defined in equation [22].
3
Defined in equation [23].
DIMORIANREVIEWwww.dimorianreview.comA PEER REVIEWED MULTIDISCIPLINARY INTERNATIONAL INDEXED E-JOURNAL
20
Vol-5, Issue-6 November-December 2018 eISSN: 2394-9163
or as:
Thus:
or as:
by combining equations [7] and [12].Equation [18] can also be written as:
A this juncture, it is quite essential to note that the ARMA (p, q) model, just like
the AR (p) and the MA (q) models; can only be employed for stationary time
series data; and yet in real life, many time series are non – stationary. For this
simple reason, ARMA models are not for describing non – stationary time series.
ARIMA models are a set of models that describe the process (for example, CPIt)
as a function of its own lags and white noise process (Box & Jenkins, 1974).
Making predicting in time series using univariate approach is best done by
employing the ARIMA models (Alnaa & Ahiakpor, 2011). A stochastic process
CPIt is referred to as an Autoregressive Integrated Moving Average (ARIMA) [p,
d, q] process if it is integrated of order “d” [I (d)] and the “d” times differenced
process has an ARMA (p, q) representation. If the sequence ∆dCPIt satisfies and
ARMA (p, q) process; then the sequence of CPIt also satisfies the ARIMA (p, d, q)
process such that:
In financial time series, it usually makes sense to take into account a model that
describes how the varience of the errors evolves and such a model is non –
other – than the ARCH model. The intuition behind ARCH family type models is
that in reality, especially in financial time series; it is rare that the varience of
the errors will be constant over time and for that reason, it is quite noble to
consider models that do not assume that the varience is constant. To expagorate
the simple intuition behind the ARCH model, I begin by defining the conditional
varience of a random variable, :
assuming thatequation [3] also holds water in this case, such that:
Equation [26] implies that the conditional varience of a zero mean normally
distributed random variable μt is equal to the conditional expected value of the
square of μt.
= + ………………………………………………………………………………………………….. [27]
Equation [27] is referred to as an ARCH (1) model since the conditional varience
depends only on one lagged squared error. This equation cannot be seen as a
complete model just because we haven’t taken into account the conditional
mean. The conditional mean, in this case; describes how the dependent variable,
CPIt; varies over time. As noted by Nyoni (2018k); there is no rule of thumb on
how to specify the conditional mean equation; actually it takes any form deemed
adequate by the researcher. Therefore, the complete model consists of both the
conditional mean equation and the ARCH specification as illustrated by Nyoni
(2018k). Equation [27] can be generalized to case where the error variance
depends on p lags of squared errors as follows:
= + +
…………...................................................................................…………….. [29]
is the simplest and yet most important case of a GARCH process, the GARCH (1,
1) model; where is the conditional varience, is the constant, is the
information about the previous period volatility, and is the fitted varience
from the model during the previous period. From equation [29], we know that:
such that:
= +( + ) + - …………………………..…………………………………………….. [31]
which is clearly an ARMA (1, 1) model; this clearly implies that indeed, a GARCH
model can be expressed as an ARMA process of squared residuals. In this case:
= - [ ] …………………………………………………..……………………………………………… [32]
is the stochastic term. Given equation [31], we can deduce that the stationarity
of the GARCH (1, 1) model requires:
+ ˂1 …………………………………………………………………………………………………………… [33]
= + + …………………………………………………………….…..………………………… [34]
so that:
= ………………………………………………………………………………………………………… [35]
For this unconditional varience to exist, equation [33] must be correct and for it
to be positive, then:
˃0 ……………………………………………….……………………………………….………………………. [36]
Equation [29] can be generalized into a GARCH (p, q) model where the current
conditional varience is parameterized to depend upon p lags of the squared error
and q lags of the conditional varience as follows:
where φ(L) and λ(L) denote the AR and MA polynomials respectively, with:
and:
or as:
= + + ………………………………...……………………………………….. [41]
+ ˂1 ………………………………………………...……………………………………………. [42]
E( )= ………………………………………………………………………..………… [45]
Conditions [33] and [42] intuitively imply the same thing. In many financial time
series, these conditions are close to unity; implying persistant volatility. If:
+ =1 …………………………………………..…………………………………………………………….. [47]
or more generally:
+ =1 ……………………………………………………………….…..……………………….. [48]
or simply:
it implies that the resulting process is not covariance stationary. Hence the term
Integrated GARCH or IGARCH model, pointing to the fact that current
information remains vital when forecasting the volatility for all horizons.
Guided by our diagnostic tests (see tables 1 – 10), we specify the following
models:
……………………………………………………. [50]
……………………………………….………………………………. [51]
The appropriate equations for the mean and varience were specified as follows:
= + + …………………………………………….………………………………………… [53]
where:
0 ……………………………………………………………….……………………………………..…….[54]
The first step towards model selection is to difference the series in order to
achieve stationarity. Once this process is over, the researcher will then examine
the correlogram in order to decide on the appropriate orders of the AR and MA
components. It is important to highlight the fact that this procedure (of choosing
the AR and MA components) is biased towards the use of personal judgement
because there are no clear – cut rules on how to decide on the appropriate AR
and MA components. Therefore, experience plays a pivotal role in this regard.
The next step is the estimation of the tentative model, after which diagnostic
testing shall follow. Diagnostic checking is usually done by generating the set of
residuals and testing whether they satisfy the characteristics of awhite noise
process. If not, there would be need for model re – specification and repetition
of the same process; this time from the second stage. The process may go on
and on until an appropriate model is identified (Nyoni, 2018i).
4
This technique was developed by Box & Jenkins (1970; 1974). For a more
simplified (and yet detailed) explanation of this technique, see Maddala (1992).
5
Retail price information is gathered on each type of product, and then weighted
according to its importance in overall consumer spending, to construct the CPI.
Monthly or annual changes in the CPI provide a good measure of the rate of
consumer price inflation (Stanford, 2008).
However, measuring CPI inflation is a tricky and at times controversial task.
Some economists believe that CPI measures actually overstate the true value of
inflation, because they do not take into account improvements in the quality of
goods and services which may offset, in some cases, some of the rise in their
prices. Other economists believe that CPI inflation may understate true inflation,
due to inadequate allowances for changes in the cost of housing and other hard
– to – estimate prices (Stanford, 2008).
DIMORIANREVIEWwww.dimorianreview.comA PEER REVIEWED MULTIDISCIPLINARY INTERNATIONAL INDEXED E-JOURNAL
25
Vol-5, Issue-6 November-December 2018 eISSN: 2394-9163
level paid by consumers for the various goods and services they
purchase.According to Jacobs et al (2014), CPI is a measure of the price of a
fixed basket of goods and services, with unchanged composition and quality
overtime.Habimana et al (2016) defines CPI as a measure of the average change
over time in the price of consumer items, goods and services that households
buy for day to day living. This study relies on inflation that was measured in
terms of CPI.
We can examine whether a series is stationary or not by analyzing the time plot
of that particular series. In this regard, a time plot of the CPI series is shown
below:
Figure 1
The above graph shows that CPI is not stationary since it is trending upwards
over the period under study. The implication is that the mean of CPI is changing
over time and hence we can safely conclude that the variance of CPI is not
constant over time.
Figure 2
The correlogram above confirms our analysis from the observation of the time
series plot of CPI. The autocorrelation coefficients are very high particulary for
the first 6 lags and is very typical in non – stationary time series data.
The Augmented Dickey Fuller (ADF test) was used to check the stationarity of
the CPI series. The general ADF test is done by running the following regression
equation:
For a purely white noise process, the autocorrelations at various lags hover
around zero (Gujarati, 2004).
Figure 3
Figure 4
While figure 4 doesn’t make it clear whether our series is now stationary or not,
tables 7 – 9 confirm that the CPI series became stationary after taking second
differences.
In this study, ARCH / GARCH effects were tested using the Langrange Multiplier
(LM) test as briefly described here: run the mean equation given by equation
[52] and save the residuals. Square the residuals and regress then on “p” own
lags to test for ARCH effects of order “p”. From this procedure, obtain and
save it. The test statistic, T (number of observations multiplied by ) follows a
distribution and the null and alternative hypotheses are:
In this research paper, the ARCH / GARCH effects test was done and the results
are shown below:
While there are a number of model evaluation criterion in time series modeling
and forecasting, for example; Mean Error (ME), Root Mean Square Error (RMSE),
Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE); this
study will only be restricted to the most commonly used and highly celebrated
criterion, that is; the Akaike ‘s Information Criteria (AIC) and the Theil’s U in
order to select the best models (in terms of parsimony [AIC] and forecast
accuracy [Theil’s U]) to be finally presented in this study. A model with a lower
AIC value is better than the one with a higher AIC value. Theil’s U, as noted by
Nyoni (2018l); must lie between 0 and 1, of which the closer it is to 0, the better
the forecast method.
Tables 12, 13 and 14 indicate that the residuals of the ARIMA (2, 2, 1) model
are stationary and bear the characteristics of a white – noise process.
The figure above shows that the ARIMA (2, 2, 1) model is stable since the
corresponding inverse roots of the characteristic polynomials are in the unit
circle.
The figure above indicates that the ARIMA (1, 2, 0) model is also stable since
the corresponding inverse root of the characteristic polynomial is in the unit
circle.
As shown in table 18 above, the mean is positive. The large difference between
the maximum and the minimum confirms the existence of an upward trend in
the CPI time series. The skewness is 1.5078 and the most striking feature is that
it is positive, implying that the CPI series has a long right tail and is non –
symmetric. The rule of thumb for kurtosis is that it should be around 3 for
normally distributed variables as reiterated by Nyoni & Bonga (2017) and yet in
this analysis, kurtosis has been found to be 1.0887. Therefore, the CPI series
not normally distributed.
Table 19
………………………………………….. [56]
……………………………………………………………..………………….. [57]
P: (0.00)
S. E: (0.12)
………………………………………………………….……………..………. [58]
P: (0.64) (0.00)
S. E (0.05) (0.00)
………………………………………………………..……….. [59]
6
The *, ** and *** means significant at 10%, 5% and 1% levels of
significance; respectively.
DIMORIANREVIEWwww.dimorianreview.comA PEER REVIEWED MULTIDISCIPLINARY INTERNATIONAL INDEXED E-JOURNAL
34
Vol-5, Issue-6 November-December 2018 eISSN: 2394-9163
ɑ1+β1 1
Since the positivity constraints specified in equations [54] are met, we safely
conclude that the estimated AR (1) – GARCH (1, 1) model is adequate. As
theoretically anticipated, and are greater than zero and is positive to
ensure that the conditional varience, is non – negative; thus the positivity
constraint of the GARCH model is not violated. Both and are positive and
statistically significant at 1% level of significance. is positive and statistically
significant, implying that strong GARCH effects cannot be ruled out. Actually, a
1% increase in previous period volatility leads to nearly 0.62% increase in
current volatility of CPI in Kenya. Since is positive and significant and indeed
less than one, it implies that the effect of old news on volatility is quite
important. In fact, an increase of 1% in previous period varience will lead to
about 0.38% increase in current volatility of CPI in Kenya. Since our empirical
model indicates that equation [47] has been confirmed, we conclude that the
estimated GARCH model is not covariance stationary; it is indeed strictly
stationary. Therefore, the estimated GARCH model is actually an IGARCH
model7. Thus current information remains essential for the forecasts of the
conditional variances for all horizons. Our findings, in this regard; are not quite
unique simply because they are similar to the findings by Moroke & Luthuli
(2015) and Nyoni (2018k).
7
The implications of IGARCH models are “too strong”. Therefore, there is need
for further research to consider fractionally integrated models such as the
Fractionally Integrated GARCH (FIGARCH) model, Fractionally Integrated
Exponential GARCH (FIEGARCH) model as well as the Fractionally Integrated
APARCH (FIAPARCH) model amongst others; when modeling and forecasting
inflation in Kenya.
DIMORIANREVIEWwww.dimorianreview.comA PEER REVIEWED MULTIDISCIPLINARY INTERNATIONAL INDEXED E-JOURNAL
35
Vol-5, Issue-6 November-December 2018 eISSN: 2394-9163
The graphs 7, 8 and 9 (with a forecast range of 10 years, that is; 2018 – 2028)
indicate that Consumer Price Index in Kenya is likely to progress in an upward
trend, implying that inflation is likely to continue rising sharply in Kenya. The
most striking characteristic of the figures 7, 8 and 9 is that they strongly concur
in their forecasts; that inflation in Kenya is indeed expected to go up, of course;
as long as no appropriate counter – cyclical policy measures are taken.
6.0 Conclusion
Monetary policy is more effective when it is forward looking (Svensson, 2005;
Faust & Wright, 2013). Maintenance of price stability continues to be one of the
main objectives of monetary policy for most countries in the world today (Nyoni
& Bonga, 2018a) and Kenya is not an exception. Achieving and maintaining price
stability will be efficient and effective the better we understand the causes of
inflation and the dynamics of how it evolves (Kohn, 2005). Therefore, inflation
forecasting is not unimportant in light of both fiscal and monetary policy making.
In Kenya, just like in any other country; inflation forecasts are envisioned to
enable the CBK to possess some control over future inflation. In this study, we
employ both ARIMA and GARCH techniques in order to model and forecast
inflation in Kenya.
References