Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

COMPARATIVE STUDY OF HOLT-WINTERS

TRIPLE EXPONENTIAL SMOOTHING AND
SEASONAL ARIMA: FORECASTING SHORT TERM
SEASONAL CAR SALES IN SOUTH AFRICA
Katleho Daniel Makatjane*, Ntebogang Dinah Moroke*
* North West University, P/ Bag X2046, Mmabatho, 2735, South Africa

Abstract
In this paper, both Seasonal ARIMA and Holt-Winters models are developed to predict the
monthly car sales in South Africa using data for the period of January 1994 to December 2013.
The purpose of this study is to choose an optimal model suited for the sector. The three error
metrics; mean absolute error, mean absolute percentage error and root mean square error were
used in making such a choice. Upon realizing that the three forecast errors could not provide
concrete basis to make conclusion, the power test was calculated for each model proving Holt-
Winters to having about 0.3% more predictive power. Empirical results also indicate that Holt-
Winters model produced more precise short-term seasonal forecasts. The findings also revealed
a structural break in April 2009, implying that the car industry was significantly affected by the
2008 and 2009 US financial crisis.

Keywords: SARIMA, Holt-Winter’ Triple Exponential Smoothing, Short-term Forecasts

1. INTRODUCTION sales in South Africa outlined. The second section

presents data and material used. Section 3 presents
Mostly, not tempered time series in global empirical analysis and results and lastly section 4
economies possess non-stationary properties and presents findings and conclusion.
are defined according to different variations such as
trends, irregular, cycles, and seasonal patterns. A 1.1 Brief background of car sales in South Africa
wide range of literature provides evidence about
different linear and nonlinear forecasting time series Automobile ownership has importantly increased
models. Some linear models such as simple linear over the world in the past two decades Shahabuddin
regression model estimation, particularly in the (2009). South Africa (SA) is no exception to this. The
context of time series modeling may give misleading country has experienced its car ownership increase
results about output-input variables nexus. from 6 million in 2000 to more than 10 million in
Moreover, conceivable issues to this may include for 2013. Prior to the season of democracy in 1994,
instance; (1) feedback from the output to the input private cars were only taken as luxury
series, (2) omitted time-lagged input term, (3) an transportation equipment on the roads of SA.
autocorrelated aggravation series and (4) basic Recently, the significance of owning a car is un-
autocorrelation patterns that are been shared by debatable globally. According to Sean et al. (2003),
variables that could create spurious relationships having a car in the United States is a second priority
Moroke (2015). with a house given the first preference. The opposite
In this study, two linear models such as the happened in the context of SA.
Holt-Winters (HW) and Seasonal Autoregressive Abu-Eisheh and Mannering (2002) emphasised
Integrated Moving Averages (SARIMA) models are that the country’s transportation infrastructure
employed to model and forecast car sales in South development is imposed by significant automobile
Africa. In the main, the study evaluates the ability of demand in travelling trends and tourism. Since
these models to handle the short-run trend with 2008, SA has invested its resources in the
seasonal components. The study further sought to development of transportation infrastructure. These
determine the model with more predictive power energetic actions significantly contribute to
and one which produces less forecasting errors. expansion of economy and creation of employment.
Modelling monthly vehicle demand is About 6% of SA gross domestic product (GDP) is
important as it provides short-term forecasts which owned by the car industry and narration for almost
assist car industries in dispatching of vehicles 12% of exports in manufacturing. Industry Export
production and this will guide policy makers on the Council, (2010). Moreover, the significance of this
demand of cars and the budget that the SA sector to the country’s economic growth cannot be
government should invest on transportation underestimated.
infrastructure. This empirical analysis is structured Based on OICA statistics, total production of
in to four sections. Firstly a brief background of car automobiles in SA had increased tremendously at

71
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

the rate of 70% for the last 13 years (1999-2013). Hannan-Quin craterial (HQC) value and 𝜀𝑡 is the error
With this being said, Chifurira et al. (2014), term.𝛽 ` 𝑠 and 𝛿are model bounds. Depending on the
developed a Johansen Cointegration and causality findings, the intercept, and intercept +trend may be
test model between SA inflation rate and new vehicle included in the model. The ADF test is defined as:
sales. This follows the results from past literature
like that of Sivak and Tsimhoni (2008), Sturgeon and 𝛾̂
Van Biesebroeck (2010) and Mimovic (2012) 𝜏= ~𝑡 , 𝑛 − 𝑝
̂ 𝛼 (5)
𝑠𝑒(𝛾)
indicating that there is a long-run relationship
between car sales and macroeconomic variables. Where the ADF test statistic is 𝜏 and 𝛾̂ is the
More equally, Pîrvu and Necşulescu (2013), process root coefficient. If the observed absolute
established a non-linear regression modelling to 𝜏 value is greater than the critical value, no simple
determine the factors that influence the decision to differencing is required since the series has been
buy a new personal vehicle. rendered stationary.

2. DATA AND MATERIALS 2.2 Material used

Data description Holt-Winters Model
The study uses a monthly car sales retrieved from Methods denoted generally as exponential
Quantec database. The series covers the period of smoothing are exceptionally well known in down to
January 1994 to December 2013 and consists of 240 earth time series smoothing and forecasting. These
observations. The data is used in its real methods are single recursive systems making such
denominations. To protect the assumption of methods simple to actualize and exceedingly and
normality, Moroke (2015) advices on the use of a computationally proficient. According to Cipra and
large sample size data. In order to stabilize the Hanzák (2008), extensions of smoothing methods to
variance factor of the series, Sadowski (2010) put the case of irregular time series analysis have
forward the log transformation as an optimal generously been presented in the past. Reference on
procedure with the standard deviation that increases the application of these methods can be made to
linearly with the mean of the series. This Cipra et al. (1995) and Cipra and Hanzák (2008).
transformation as highlighted by Montgomery et al. Chatfield and Yar (1988a) viewed the Holt-
(2015) follows the form: Winters model as a variation of exponential
smoothing which is straightforward, yet by large
(𝑋𝑡 − 𝑋𝑡−1 ) practices, is admirable. This is a special short-term
𝑌𝑡 = (1)
𝑋𝑡−1 forecast model in demand and sales time series.
Literature on variables exhibiting seasonal trends
According to Bruce et al. (2005), pre- through the use of exponentially weighted moving
differencing transformation should sometimes be average (EWMA) methods by Holt (2004) reports that
employed so as to stabilize the properties of the a time series either has a trend additive,
series. Statistical Analysis Software (SAS) version 9.3 multiplicative or multiplicative error structure
is used for data analysis. components. In dealing with seasonal and trend
The two methods used are built on the basis of forecast, the EWMA according to literature is
Box-Jenkins methodology which allows only a reported to be the best model. The smoothing
stationary series before model estimation. For linear equations of Holt-Winters method have two
time series modelling, the Augmented Dickey-Fuller approaches. The additive and multiplicative aproach
(ADF) unit root test as recommended by Mushtaq is as defined as follows.
(2011) is used. This test in linear regression form is
written as: Multiplicative Holt-Winters Method
ADF equation with no intercept and no trend:
𝑝 The Level Equation:
∆𝑋𝑡 = 𝜌𝑋𝑡−1 + ∑ 𝛿𝑖 ∆𝑋𝑡−𝑖 + 𝜖𝑡 (2) 𝑋𝑡
𝑖=1 ℓ𝑡 = 𝛼 ( ) + (1 − 𝛼)(ℓ 𝑇−1 + 𝛽𝑡−1 ), (6)
𝑠𝑛𝑡−𝑙
ADF equation with intercept:
The Growth Equation:
𝑝

(3) 𝛽𝑡 = 𝛾(ℓ𝑡 − ℓ𝑡−1 ) + (1 − 𝛾)𝛽𝑡−1 , (7)

∆𝑋𝑡 = 𝛽0 + 𝜌𝑋𝑡−1 + ∑ 𝛿𝑖 ∆𝑋𝑡−𝑖 + 𝜖𝑡
𝑖=1
The Seasonal Factors Equation:
ADF equation with intercept plus trend: 𝑋
𝑆𝑛𝑡 = 𝛿 ( 𝑡 ) + (1 − 𝛿)𝑆𝑛𝑡−𝐿 , (8)
𝑝 ℓ𝑡

∆𝑋𝑡 = 𝛽0 + 𝛽1 𝑡 + 𝜌𝑋𝑡−1 + ∑ 𝛿𝑖 ∆𝑋𝑡−𝑖 + 𝜖𝑡 (4) where 𝛼, 𝛾, 𝑎𝑛𝑑 𝛿 are the smoothing constants
𝑖=1 between 0 and 1, ℓ𝑡−1 𝑎𝑛𝑑 𝛽𝑡−1 are estimates in time
period 𝑡 − 1 for level and growth equation, and 𝑆𝑛𝑡−1
 is a differencing operator, t is time drift; p is the seasonal factor estimate in time period 𝑡 − 𝐿.
denotes the selected maximum lag based on the Note that, the seasonal length adds up to the length
minimum criteria such as Aikaike’s information of the season, that is, for monthly seasonal data
criteria (AIC), Schwatz Bayesian crateria (SBC) or 𝑆𝑛 = 12 for quarterly data 𝑆𝑛 = 4 and so on and

72
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

forth. The trend component 𝛽𝑡 if deemed Taylor (2003) when predicting electricity demand.
unnecessary is deleted from the model yielding a The wellspring of this correlation may be because of
model with damped trend as: elements of the series which expressly do not take
The Level Equation: into consideration the details of the states. For
instance, the yearly seasonal effects might affect the
𝑋𝑡 series and the constrained sample size implies that
ℓ𝑡 = 𝛼 ( ) + (1 − 𝛼)(ℓ 𝑇−1 + 𝜙𝛽𝑡−1 ) (9)
𝑠𝑛𝑡−𝑙 it cannot be unequivocally modelled. This is the
discussion of De Livera et al. (2011). It was
The Growth Equation: previously suggested that all exponential smoothing
methods be regarded as a special case for ARIMA
𝛽𝑡 = 𝛾(ℓ𝑡 − ℓ𝑡−1 ) + (1 − 𝛾)𝜙𝛽𝑡−𝐿 (10) models, but this view has been ignored in recent
years. There is no distinct comparison between the
The Season Factors Equation: additive seasonal Holt-Winters model and ARIMA
because the former is classified as a complicated
𝑆𝑛𝑡 = 𝛿(𝑋𝑡 /ℓ𝑡 ) + (1 − 𝛿)𝑆𝑛𝑡−𝐿 (11) ARIMA model (Taylor, 2003).
A point forecast made in time period T for
The K-step forecast estimator of EWMA method X T  (T ) is:
is defined by the following equation:

𝑋̂𝑡 (𝑘) = 𝛼𝑋𝑡 + (1 − 𝛼)𝑋̂𝑡−1 (𝑘) (12) 𝑋̂𝑡+𝜏 (𝑇) = (ℓ𝑡 + 𝜙𝛽𝑡 + 𝜙 2 𝛽𝑡 + ⋯ + 𝜙 𝜏 𝛽𝑡 )𝑆𝑛𝑡+𝜏−𝐿 (15)

where 𝛼 is the smoothing parameter that lies Seasonal Autoregressive Integrated Moving
between 0 < 𝛼 < 1 with 𝜀𝑡 = 𝑋𝑡 − 𝑋̂𝑡−1 (𝑘) being a k- Average
step-ahead forecast error at time t. ARIMA models have been pioneered by Box and
Holt (2004); recommended this approach when Jenkins (1976). These models are intended for the
time series is in the form of a trend and irregularity. forecasting of traffic flow data and have since been
A trend is regarded as a long-term change in the successfully used. The general SARIMA model
mean level per unit time. On the off chance that following Box et al. (2011) is:
trend is thought to be linear, it is vital to recognize a
worldwide linear trend of the structure: Φ(𝐿)Φ𝑠 (𝐿𝑠 )(1 − 𝐿) 𝑠 ((1 − 𝐿)𝐷 𝑋𝑡 = Θ(𝐿)Θ𝑠 (𝐿𝑠 )𝜖𝑡 , (16)

𝜇𝑡 = 𝛼 + 𝛽𝑡 . (13) with 𝜀𝑡 ~𝑖𝑖𝑑(0, 𝜎𝑡2 ), and 𝑆 being the seasonal

length as just like in Holt-Winters model. As a result,
If 𝛼 and 𝛽are estimated parameters, then the 𝑋𝑡~ARIMA (𝑝, 𝑑, 𝑞)(𝑃, 𝐷, 𝑄)𝑠.
linear trend is: ARIMA model has been perfectly employed to a
space and time factors to forecast a space-time
𝜇𝑡 = 𝛼𝑡 + 𝛽𝑡 𝑡 (14) stationary traffic flow by both Kamarianakis and
Prastacos (2005) and Ding et al. (2011). Emphasized
where 𝛼𝑡 𝑎𝑛𝑑 𝛽𝑡 change slowly through time in a by DA VEIGA et al. (2014), literature on ARIMA
random way and the quantity β or 𝛽𝑡 is a trend. model is alluring due to its theoretical properties
With respect to seasonality, the principle and some supporting evidence from various
refinement is between the additive seasonality and empirical. The drawbacks of the ARIMA model are
multiplicative seasonal elements (Holt, 2004). The identified as its pure direction to focus on the past
latter being appropriate when the magnitude of mean values and inability to capture the fast
seasonal variation is relative to the nearby mean. growing variation within the inter-urban traffic flow
Nonetheless, Chatfield and Yar (1988b), emphasized Hong et al. (2011). Any forecasting technique
that there is some sort of relationship between Holt- includes two stages such as the analysis of time
Winter methodology and other procedures series and the choice of forecasting model that best
specifically Box-Jenkins methodology (example, Box fits the data set. ARIMA model is utilized in a
et al., 2011) and the use of state-space or structural comparative grouping of analysis and selection by
models. decomposition methods and regression.
According to literature, simple exponential The expansion of the ARIMA model for traffic
smoothing is approximately 𝐴𝑅𝐼𝑀𝐴 (0, 1, 1) model. A flow has recently been exploited by Williams and
counterpart double exponential smoothing also Hoel (2003). The authors applied ARIMA model with
known as two-parameter (non-seasonal) model is seasonal peak or non-peak periods. The findings of
said to be a 𝐴𝑅𝐼𝑀𝐴 (0, 2, 2) model. All exponential their study revealed a significant heuristic
smoothing methods need some estimation of forecasting accuracy by the model. These new
smoothing parameters which is either 𝛼 𝑜𝑟 𝛾 Hilas et discoveries reassure authors to utilize SARIMA
al. (2006) highlighted that the minimization of the model. Moroke (2014) also used SARIMA in
mean square error is the common method of forecasting the SA household debts. The results of
estimating the parameters and this is normally done this study reported this model to be robust in
through the grid search method. producing the forecasts of this sector. To capture
The error process 𝑑𝑡 is said to be free from the seasonality in time series, there is a strong appeal to
serial correlation when estimating with smoothing select a more flexible forecasting model and this
models. More often than not, this might not be the task is fulfilled with SARIMA and the Holt-Winters
case. Chatfield and Yar (1988a) used Holt-Winters methods. Chikobvu and Sigauke (2012) and Ghosh
multiplicative algorithm for seasonal effects and (2008) also used SARIMA model in producing short-
found the error term to be an autoregressive of term forecasts of electricity successfully.
order one (AR (1)). Similar findings were reported by

73
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

Structural change test when used on samples in excess of 50 observations.

From the power computations, the JB test is found
In order to identify and encounter for the structural to have a large empirical alpha test of normality for
change in the sale of cars in SA, the Chow test is both small and large samples hence it is the best
estimated as to offer the classical possibility of over the other normality tests. The JB test is
structural change. The test is estimated as: calculated using the formula:

(𝜀𝑡′ 𝜀𝑡 − 𝜀𝑡1
′ ′
𝜀𝑡1 − 𝜀𝑡2 𝜀𝑡2 )/𝑘 𝑛−𝑘 2 1
𝐶ℎ𝑜𝑤 = (17) 𝐽𝐵 = (𝑆 + (𝐾 − 3)2 ) ~𝜒 2 , 2𝑑𝑓 (20)
′ ′
𝜀𝑡1 𝜀𝑡1 + 𝜀𝑡2 𝜀𝑡2 )/(𝑛1 + 𝑛2 − 2𝑘 6 4

where 𝑆 is the skewness, 𝐾 is the number of

t is the residual vector from the entire regressors from the regression model, 𝑛 is the
sample size and 2𝑑𝑓 is the number of degrees of
regression data set, 𝑘 and 𝑛1 + 𝑛2 − 2𝑘 are the
freedom. The test follows a chi-square distribution
number of degrees of freedom, 𝜀1 𝑎𝑛𝑑 𝜀2 are the
with 3 degrees of freedom for sample size of 2000
residual from the subset regressions. The subset
and above. But when the sample is less than 2000,
regressions are as follows;
the JB test follows a normal cumulative distribution
(NCD). The tested hypothesis is:
𝑌𝑡1 = 𝛽𝑡1 𝑋𝑡1 + 𝜀𝑡2 , 𝑛1 (18a)
H 0 : E ( t )  0
𝑌𝑡2 = 𝛽𝑡2 𝑋𝑡2 + 𝜀𝑡2 , 𝑛2 (18b)
H a : E ( t )  0
𝑛1 and 𝑛2 are number of observations. The null hypothesis is rejected if the calculated
The main focal point of the test is to test the probability value of the JB static is less than an
stability of a relationship between a response observed probability value or if the calculated JB
variable and the explanatory regressor. If there is no statistic is greater than the critical value obtained
structural change, the estimated residuals from the from chi-square distribution with two degrees of
regression using the entire data is expected not to freedom.
differ from the combined residuals from the two
regressions using each subset of the data. However, Serial correlation
a large difference between the sets of residuals
indicates that there has been a break in the data at While the Durbin-Watson test is formulated with the
the specified period. 𝐴𝑅(1) alternative hypothesis error; it should have
some power in detecting other forms of serial
Information criterion for model selection between correlation provided 𝐸[𝜀𝑡 𝜀𝑡−1 ] ≠ 0 under the
the candidate models alternative hypothesis. Still, there are more powerful
tests for high-order serial correlation that involves
Model selection is an important issue in almost any high-order autocorrelation estimators. For high
practical data analysis; the model might have a large order test, the Breusch-Godfrey test is used in this
R2 but will give spurious results. The main objective study. Suppose the error terms are 𝐴𝑅(𝑝) for 𝑝 > 1
of the current study is to select the best model by i.e.
the use of Schwarz Bayesian information criterion
(SBC) for both Holt-Winters and SARIMA model. Note 𝜀𝑡 = +𝜌1 𝜖𝑡−1 + ⋯ + 𝜌𝑝 𝜀𝑡−𝑝 + 𝜈𝑡 , (21)
that the model with the smallest SBIC is preferred
and the estimation of the SBIC is based on the and 𝜈𝑡 ~𝑖. 𝑖. 𝑑(0, 𝜎 2 )
likelihood function and it was developed by Schwarz The hypothesis here is defined as:
(1978) and introduced it to follow the form:
𝐻0 : 𝜌1 = 𝜌2 = ⋯ = 𝜌𝑚
SBC = −2[ln 𝐿̂ + 𝑘 ∗ ln(𝑛)] (19) 𝐻𝑎 : 𝜌𝑚 ≠ 0

where n is the sample size and k is the number Equation 20 is a Q-statistic of squared residuals
of parameters to be estimated and 𝐿̂ is the likelihood which is given by:
function of the estimated model (M) which is 𝑝
𝐿̂ = 𝑝(𝑥|𝜃̂, 𝑀) and x is the observed data and 𝜃 is the 𝜌𝑘2
𝑄𝑚 = 𝑛(𝑛 + 2) ∑ ~𝜒 2 , 𝑚 − 𝑝 (21)
parameters of the estimated model. 𝑛 − 𝑘 𝛼,
𝑘=1

Assumptions and model diagnostics Heteroscedasticity

This section discusses the tests for the assumptions To test for heteroscedasticity in 𝜀𝑡 , the ARCH test is
such as normality, serial correlation and
employed. The test statistic is an extended high
heteroscedasticity in that respect.
order 𝐴𝑅𝐶𝐻𝑝 effects which is presented as:
Normality 2
𝑉𝑎𝑟(𝜀𝑡 ) = 𝛾0 + 𝛾1 𝜀𝑡−1 + ⋯ + 𝛾𝑝 (23)
Jarque-Bera (JB) test is used in this study to test the
Therefore the Lagrange Multiplier (LM) test for
speculation about the fact that a given sample 𝑋𝑠 is a
heteroscedasticity is:
specimen from a normal distribution. An also the
estimated residuals for each model are normally
𝐿𝑀 𝑆𝑡𝑎𝑡𝑖𝑠𝑡𝑖𝑐 = (𝑛 − 𝑝)𝑅2 ~𝜒𝑝2 (24)
distributed. The JB test of normality performs better

74
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

𝑛
with 𝑝 being the number of estimated 2
parameters, n is the sample size and 𝑅2 is the 𝑀𝑆𝐸 = ∑[𝑋𝑡 − 𝑋̂𝑡 ] (25)
adjusted 𝑅2 which comes from the squared 𝑡=1
regression model in (21). Hence the tested 𝑛
hypothesis is 1
H0 : Var(εt ) = σ2t 𝑀𝐴𝐸 = ∑|𝑋𝑡 − 𝑋̂𝑡 | (26)
𝑛
𝑡=1
H1 : Var(εt ) ≠ σ2t
Here, the test rejects the null hypothesis if the 1
𝑀𝐴𝑃𝐸 = ∑𝑛𝑡=1|𝑋𝑡 − 𝑋̂𝑡 |*100 (27)
LM test is greater than critical value of 𝜒 2 , 𝐿 − 𝑛
1 𝑑𝑓 and conclude that the error term is constant
over time. 3. EMPIRICAL ANALYSIS

Forecasting performance test This section provides and discusses the preliminary
and primary analyses results.
To check forecasting performance of each model,
the performance error metrics are recommended for 3.1 Preliminary results
evaluating models. In order to select the appropriate
model between the two linear models namely Holt- In this section the preliminary data analyses are
Winters and SARIMA models, three error metrics, conducted with the purpose of assessing the
mean square error (MSE) and mean absolute error behavior of the data set. In the current study, the
(MAE) and mean absolute percentage error (MAPE) adoption of the descriptive statistics is used to
are appealed to. Given the time series, 𝑋𝑡 and provide a sound understanding of the data. Table 1
estimated series,𝑋̂𝑡 , the three error metrics are presents the summary statistics from the SA car
defined below: sales data.

Table 1. Exploratory data analysis

Variable Car Sales

Mean 49374.68
Std Deviation 18512.61
Skewness 0.39303
JB 0.5818
P-Value 0.6178

The mean value of car sales in Table 1 revealed industries closing down. SA also suffered economic
that, on average the SA economy is selling 49 375 recession, hence a dip in 2009.
cars monthly. This implies that in SA context, the The cause of this intense change is the increase
whole period of 1994-2013 the number of cars in unemployment and poverty in the whole world
produced and sold in a month is 49375. The JB which contributed to the decline in aggregate
normality test is 0.5818, and the associated demand. According to Moroke et al. (2014), the
probability value is 0.6178 which is greater than 2007-2009 crisis had a colossal effect on economies,
0.05. This provides evidence to conclude that the with securities exchanges falling, financial
data comes from a normal distribution. institutions caving in and governments been
Next, the paper presents the results for the compelled to intercede with bailouts, while trying to
structural break as depicted in 2009 in Figure 1 and put more attention on administrative change. This
Tables 2 and 3. also brought a significant drop on the economic
growths globally. The South African Reserve Bank
3.2 Structural Break Test (SARB) 2010 quarterly report uncovers that South
Africa's GDP was 15.3% in 2009. Currently the rate of
In univariate timeseries analysis, the overlay plots economic growth in SA is at 2% as per annual
are normally adopted to check the behaviour of the bulletin from the SARB.
data. Figure 1 is the plot of monthly car sales from
1994 to 2013 in SA. The figure shows a roughly
increasing seasonal trend. This implies that the
series of car sales is nonstationary. Generally, car
industry in the country was doing well with some
time epochs during some seasons. In the 184th
observation, i.e. monthly sales in April 2009, there
was a break from the sales of cars in SA as shown by
a profound dip. This period marks a numerical drop
from 53,000 cars in March to 38,200 cars in April. It
should be noted that, most of the countries suffered
the spill-over effects of US financial crisis which
occured between 2007-2009. These effects started
hiting most economies after 2009 and during that
time most financial sectors of different countries
suffered the effects causing the slowing down in
production, people being retrenched and most

75
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

Figure 1. Total Monthly Car Sales

90000

80000

70000

60000

50000

40000

30000

20000

10000
0 8 1 2 3 4 4 5 6 7 8 8 9 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2
6 4 2 0 8 6 4 2 0 8 6 0 1 2 2 3 4 5 6 6 7 8 9 0 0 1 2 3 4
4 2 0 8 6 4 2 0 8 6 4 2 0 8 6 4 2 0
PLOT Car Sales in South Africa

Table 2. Chow structural change test

Test Break Point Num DF Den DF F Value Prob

Chow 184 13 214 4.34 **

Notes *** significant @ 10% ** significant at 5%, *significant at 1%; N not significant

The Chow test in Table 2 supports a significant This implies that this interaction is precipitous with
structural break on the 184th observation associated a permanent duration and this means that there is a
with April 2009 also vissible in Figure 1. The dummy structural break on April 2009. Chow test also
interaction variable (Dum1) in Table 3 is also confirms that there was a significant drop of cars
statistically siginificant at 5% level of significance. sales in SA in April 2009.

Table 3. The Piecewise Regression Parameter Estimates

Variable Coefficient Std. Error t-Statistic Prob.

C 19635.99 1327.861 14.78768 **
T 263.3427 12.50641 21.05661 **
DUM1 -7417.515 2048.603 -3.620767 **

Notes *** significant @ 10% ** significant at 5%, *significant at 1%; N not significant

The results in Table 3 indicates that there is a All of this confirmed Naudé (2009) warnings about
decline of about 7417 in car sales per month. the spillover effects of the financial crisis, especially
Though SA was not deeply affected by the 2007- to Africa and those countries dependent on the US
2009 financial crisis, the industries within the for trade.
country were knocked down by the financial To accommodate the Box-Jenkins methods
overflow. Resources got downgraded, companies proposed for this study, the series in Figure 1 is log
were shut down causing unemployment rates to differenced to help stabilize the properties of time
accelerate profusely with the overall diminishing of series. The results are shown as Figure 2 and Table
the country’s economic growth (Moroke et al., 2014). 4.

Figure 2. First Seasonal Differencing

76
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

By visual inspection, the partial autocorrelation quickly implying that the properties of the series are
function (PACF) and autocorrelations function (ACF) not time-variant. The statistical test results are
of the first log seasonal differencing of the car sales presented in Table 4.
is stationary. The spikes of these functions die

Table 4. Augmented Dickey Fuller Test

Type Lags Rho Prob Tau Prob

Zero Mean 0 -299.247 * -20.96 *
1 -636.153 * -17.74 *
2 -509.297 * -10.74 *
3 -589.248 * -8.86 *
Single Mean 0 -299.275 * -20.92 *
1 -636.514 * -17.71 *
2 -510.448 * -10.72 *
3 -593.203 * -8.84 *
Trend 0 -299.283 * -20.87 *
1 -636.665 * -17.67 *
2 -510.382 * -10.69 *
3 -593.202 * -8.82 *

Notes *** significant @ 10% ** significant at 5%, *significant at 1%; N not significant

The ADF test confirms that the series is

stationary at 5% level of significant at all lags. The
findings also show that the model with the three 3.3 Holt-Winter’s Exponential Smoothing Results
features will produce better forecasts for car sales in
SA. Primary data analysis is performed on The estimated Holt-Winters model is reported in
stationarised data containing the three features and table 5. Both multiplicative seasonal model with
the results are presented in the next sections. trend and multiplicative seasonal model are
estimated. The best model is selected by the use of
SBC which then indicates that the Multiplicative
seasonal model is the best because the reported SBC
is -1163.4 compared to the winter’s model.

Table 5. Holts-Winters Triple Exponential Smoothing Results

_MODEL_ _PARM_ _EST_ _STDERR_ _TVALUE_ _PVALUE_

WINTERS LEVEL 0.49817 0.039425 12.636 0.000
WINTERS TREND 0.001 0.013329 0.075 0.94026
WINTERS SEASON 0.06647 0.029017 2.2909 0.02285
R2 0.94871
SBC -1160.4
MULTSEASONAL LEVEL 0.5116 0.039604 12.9177 0.000
MULTSEASONAL SEASON 0.06853 0.029842 2.2964 0.022521
R2 0.94816
SBC -1163.4

𝑋̂𝑡 = (0.51160)(0.06853)𝑆𝑡 And the associated time. And, the seasonal factors are reported in table
𝑅2 = 0.94816 which means the model is significant 6 as.
because 95% of variation in car sales is explained by

Table 6. The estimated seasonalities

JAN FEB MAR APR MAY JUN

0.98963 1.0011 1.00893 0.98958 1.00128 1.00545
JUL AUG SEP OCT NOV DEC
1.00707 1.0038 0.9979 1.01053 1.00271 0.98287

The results in table 6 implies that throuhgout the al. (2015) cleared that this cyclical parttens are even
whole period of 1994-2013, there were some likely to happen in an out-of sample forecasts.
irregularities that happened within South African car
industry. The estimates are not constant throughout
the months as dipicted by Table 5. Montgomery et

77
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

3.3.1 Box-Jenkins SARIMA Results which was suggested by Stadnytska et al. (2008). The
use of the traditional way of using the PACF and ACF
Since both SARIMA and Holt-Winters method are plots is employed by then examining the behavior of
capable of capturing the short run seasonalities the two plots and then select the model based on the
within the data, with SARIMA models, the first step number of spikes outside the confidence bend of the
is to difference the seasonal and non-seasonal series plots. And finally model estimation is done through
so as to enable the selection of the best model maximum likelihood estimation method. The
among candidate models. Which is done through the parameter estimates of the SARIMA model are
use of the Bayesian Information Criterion (BIC) The reported in table 7.
model is identified by an automated procedure

Table 7. Parameter estimates of SARIMA model

Parameter Estimate Standard Error t Value Pr > |t| Lag

MA1,1 0.5138 0.05806 8.85 0.0001 1
AR1,1 0.20235 0.06619 3.06 0.0022 3
AR2,1 -0.39939 0.06264 -6.38 0.0001 12

The estimated model can be written as: McGee (2000) suggested, the estimates of the model
must be less than one to deem them sufficient and
[1 − 0.20235Φ ∗∗∗ (3)(1 + 0.39939Φ ∗∗∗ (12))]𝑋𝑡 significant. Note that the first-order pure seasonal
= [1 − 0.5138Θ ∗∗ (1)]𝜀𝑡 , differencing was established to obtain stationary
and, 𝜀𝑡 ~𝑖𝑖𝑑(0,0.12710). values of the series.
The point estimates of this model are all
significant. These outcomes are as per those of the 3.3.2 Model diagnostic check
correlational and full limit versatility theory
acquired in the preparatory analyses. As Yaffee and

Table 8. Holt-Winters and SARIMA

Holt-Winters SARIMA
Statistic Estimate Prob Statistic Estimate Prob
JB test 2.9655 N JB test 2.7298 N
Godfrey test Godfrey test
AR(1) 0.0012 N AR(1) 0.7622 N
AR(2) 3.0377 N AR(2) 5.1554 N
AR(3) 11.361 ** AR(3) 5.1562 ***
ARCH test 1.65 N ARCH test 0.002153 N

Notes *** significant @ 10% ** significant at 5%, *significant at 1%; N not significant

Table 8 presents model diagnostic tests for three error metrics discussed in Section 2 are used
both Holt-Winters and SARIMA models. At 5% level to measure the performance of each model and
of significant, both model residuals met all the results are summarized in Table 9. Some tentative
assumptions. For serial correlation, the Godfrey test conclusions are drawn from these table, which
for AR (1) and AR (2) have insignificant probability indicates that the Holt-Winters model dominates the
values, hence the null hypothesis cannot be rejected. SARIMA model.
It is concluded that there the two model residuals 1) In terms of MAE, the Holt-Winters model has
are not serially correlated. Lastly the ARCH test also achieved the smallest value which is 2414.62
reveals that both models residuals are not compared to SARIMA model.
heteroscedastic. This implies that over the period of 2) SARIMA model has performed better when it
1994-2013, the residual of the two model are comes to the MAPE statistic.it has a smallest of
constant over time. Having established that the 0.067547 compared to Holt-Winters model.
assumptions of the models are not violated, these 3) For the MSE, the Holt-Winters model produced
models are used for further analysis and the results the small error of 8957466.84 which makes it
are presented in the next section. the better model compared to SARIMA model.
Furthermore, the final conclusion is made and
3.4 Comparative Analysis found that in general the Holt-Winters model is the
one with less forecasting errors in forecasting the
The purpose of this section is to determine the car sales data in SA. This selection is due to the fact
model which best mimics the data and also that MAE and MSE of the Holt-Winters model are the
produces fewer forecasts. This will help in assisting smallest compared to the SARIMA model.
the maximum dispatching of SA’s car industries. The

78
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

Table 9. Performance model selection Criteria

Performance Criteria Holt-winters Model SARIMA Model

MAE_Ratio 2414.62 4288.78
MAPE_Ratio 0.071599 0.067547
MSE_Ratio 8957466.84 27071931.35

Another interesting results is the reported monthly seasonality. These statistics are all less for
descriptive statistics of the error forecasting Holt-Winters triple exponential smoothing than
measures in Table 10. Looking at standard those of SARIMA model which in return gives the
deviations, there is enough evidence at which is the small forecasting confidence intervals as compared
best model for capturing short-term seasonal to those of the SARIMA model. Hence Holt-Winters
components. The minimum and maximum values of model is favored over SARIMA model. Hence the
the error forecasting measures pick the Holt-Winters results of the confidence interval for both models
method as the better model for the short-term are reported in Table 11.

Table 10. Descriptive Statistics of Error Forecasting Measures

Model Variable Mean Std Dev Minimum Maximum

MAE_Ratio 2414.62 1775.77 4.4535106 8316.66
Holt-winters Model MAPE_Ratio 0.0715989 0.0544593 0.00013085 0.3380327
MSE_Ratio 8957466.8 11723587.5 19.8337571 69166839.1
MAE_Ratio 4288.78 2958.24 88.427676 13713.57
SARIMA Model MAPE_Ratio 0.0675474 0.0499125 0.0012291 0.3146396
MSE_Ratio 27071931 34783994.5 7819.45 188061871

Table 11. Confidence interval for Holt-Winters and SARIMA Models

Holt-Winters Model SARIMA Model

Date Forecast L0.95 U0.95 Forecast L0.95 U0.95
Jan-2014 11.2547 11.0412 11.4682 11.167 10.9605 11.3735
Feb-2014 11.3005 11.0558 11.5452 11.3021 11.0724 11.5317
Mar-2014 11.2862 11.0321 11.5403 11.3477 11.0971 11.5983
Apr-2014 11.257 10.9873 11.5268 11.1775 10.8893 11.4657
May-2014 11.3086 11.0269 11.5903 11.2882 10.9757 11.6006
Jun-2014 11.3009 11.0121 11.5896 11.3069 10.972 11.6419
Jul-2014 11.3417 11.047 11.6363 11.3791 11.0201 11.7381
Aug-2014 11.2999 11.0005 11.5994 11.3127 10.9326 11.6928
Sep-2014 11.22 10.917 11.5231 11.1923 10.7922 11.5923
Oct-2014 11.3724 11.0666 11.6782 11.3459 10.9263 11.7655
Nov-2014 11.3306 11.0226 11.6386 11.3019 10.8639 11.7399
Dec-2014 11.2638 10.9541 11.5736 11.1554 10.6998 11.6111

Table 11 present the confidence intervals of the that the smaller the standard deviation of the error,
forecasts for both Holt-Winters and SARIMA models. the less wider the interval. Hence the Holt-Winters is
From the interval estimates, the Holt-Winters selected over SARIMA.
produced small interval range for the 12 month
forecasts confirming the results reported in Table 9

Figure 2. Fitted conditional mean for car sales

90000

80000

70000
Holt-Winters model

60000

50000

40000

30000

20000

10000
J J J J J J J J J J J J J J J J J J J J J
A A A A A A A A A A A A A A A A A A A A A
N N N N N N N N N N N N N N N N N N N N N
9 9 9 9 9 9 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1
4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4
Date
PLOT Car Sales in South Africa Est. Car Sales

79
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

100000

90000

80000

70000

SARIMA MODEL
60000

50000

40000

30000

20000

10000
J J J J J J J J J J J J J J J J J J J J J
A A A A A A A A A A A A A A A A A A A A A
N N N N N N N N N N N N N N N N N N N N N
9 9 9 9 9 9 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1
4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4
Date
PLOT Car Sales in South Africa Xhat

Figure 2 presents the fitted conditional mean seasonal of the data at hand. Hence the power test
for both models. By visual examination, both models of each model through their estimated means
seems to fit the car sales data well. But between the confirms that the best and powerful model is the
two models, there is that one which is more Holt-Winters. The reported power of the Holt-
powerful than the other and the results are reported Winters in table 11 is 0.803 which is 0.3% more
in tables 9 and 10 pics the Holt-Winters model as the powerful than the counterpart SARIMA model.
most powerful to model and forecast the short-term

Table 11. Model Power Test

Model Mean Diff Actual Power

SARIMA -0.00087 0.8
HOLT-WINTERS 0.07168 0.803

3.5 Forecasting with Holt-Winters Model selected model. In this case, the selected model is
the Holt-Winters multiplicative model with damped
After selecting the best model that best fit the data trend. The expected forecasts are presented in Table
well, the short-term forecasts are produced with the 12.

Table 12. Seasonal Forecasting results with Holt-Winters model

Date Estimated Car sales

Jan 2014 68545.24
Feb 2014 77883.88
Mar 2014 84876.81
Apr 2014 68540.34
May 2014 77986.2
Jun 2014 81268.59
Jul 2014 84430.73
Aug 2014 82158.65
Sep 2014 77565.73
Oct 2014 85780.23
Nov 2014 79748.28
Dec 2014 64818.68

There is an increased car sales in SA with a dip ARIMA were used and their results were compared.
in December 2014. This also implies that the Upon realizing that the two models were adequate
demand may be higher that of previous years and as and did not violate any of the assumptions, they
a result more resources may be needed to meet the were used for further analysis. To assess the
demand. This could be good news to South Africans capability forecasting of the two models, a level of
as more jobs may be made available. equal methodology was estimated and for the
accuracy of forecasting, three measures were
4. CONCLUDING REMARKS constructed. The findings proved that Hot-Winters
triple exponential smoothing was a more powerful
This study sought to determine the model which can compared to SARIMA hence the latter was deemed
be suited to forecast car sales in South Africa. an optimal model by the study.
Monthly car sales data used was obtained from Based on these findings, several
Quantec database covering the period January 1994 recommendation are made for further studies and to
to December 2013. Two univariate models known the car industry officials. The decision in utilizing a
for producing short term forecasts such as Holt- Box–Jenkins or Holt-Winters models relies upon the
Winters triple exponential smoothing and seasonal expected use of the series which can sensibly be

80
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

thought continuous, a whiz decision would to be Transactions on Business and Economics, 11:608-
apply a Holt-Winters multiplicative approach. 614.
Despite the fact that Box–Jenkins and Holt-Winters 11. De Livera, A.M., Hyndman, R.J. & Snyder, R.D. 2011.
models have comparable forecasting ability on car Forecasting time series with complex seasonal
sales data, the latter is more adaptable for managing patterns using exponential smoothing. Journal of
distinctive data scenarios. The reported quantitative the American Statistical Association, 106
comparison between SARIMA and Holt-Winters is
(496):1513-1527.
emphatically reliant on the time series and the
12. Ding, Q.Y., Wang, X.F., Zhang, X.Y. & Sun, Z.Q. 2011.
chosen error measures. Extra assessment of both
models was established and found that Holt-Winters Forecasting traffic volume with space-time ARIMA
has more predictive power than SARIMA. For more model. (In Advanced Materials Research organised
interesting studies, a researcher can even include by Trans Tech Publ.
simulated data sets and compare the SARIMA 13. Hilas, C.S., Goudos, S.K. & Sahalos, J.N. 2006.
models and Holt-Winters models with other time Seasonal decomposition and forecasting of
series techniques, for instance, artificial neural telecommunication data: A comparative case study.
networks. On the side of policy makers, policies Technological Forecasting and Social Change, 73
regarding the car industries must be re-evaluated. (5):495-509.
Firstly, national roads should be improved as the 14. Holt, C.C. 2004. Forecasting seasonals and trends
forecasts indicated that on monthly bases, the sales by exponentially weighted moving averages.
of car are increasing over time. This will also bring
International Journal of Forecasting, 20 (1):5-10, 1.
more income to the South Africa economy through
15. Hong, W.-C., Dong, Y., Zheng, F. & Wei, S.Y. 2011.
the tourism sector as more people will be visiting SA
and as a result GDP will be boosted. Moreover, Hybrid evolutionary algorithms in a SVR traffic
future economic policy should focus more on new flow forecasting model. Applied Mathematics and
vehicle manufacturing, the sector has the potential Computation, 217 (15):6733-6747.
to grow and generate employment and more earning 16. Kamarianakis, Y. & Prastacos, P. 2005. Space–time
to South Africa. modeling of traffic flow. Computers & Geosciences,
31 (2):119-133.
REFERENCES 17. Mimovic, P. 2012. Applications of Analytical
Network Process in Forecasting Automobile Sales
1. Abu-Eisheh, S.A. & Mannering, F.L. 2002. of Fiat 500 L. Economic Horizons, 14 (3):169-179.
Forecasting automobile demand for economies in 18. Montgomery, D.C., Jennings, C.L. & Kulahci, M.
transition: A dynamic simultaneous-equation 2015. Introduction to time series analysis and
system approach. Transportation Planning and forecasting. John Wiley & Sons.
Technology, 25 (4):311-331. 19. Moroke, N.D. 2014. The robustness and accuracy of
2. Box, G.E. & Jenkins, G.M. 1976. Time series analysis, Box-Jenkins ARIMA in modeling and forecasting
control, and forecasting. San Francisco, CA: Holden household debt in South Africa. Journal of
Day, 3226 (3228):10. Economics and Behavioral Studies, 6 (9):748.
3. Box, G.E., Jenkins, G.M. & Reinsel, G.C. 2011. Time 20. Moroke, N.D., Mukuddem-Petersen, J. & Petersen, M.
series analysis: forecasting and control. Vol. 734: 2014. A Multivariate Time Series Analysis of
John Wiley & Sons. Household Debts during 2007-2009 Financial Crisis
4. Bruce, L., Richard, T. & Anne, B. 2005. Forecasting in South Africa: A Vector Error Correction
Time Series and Regression. Thomson Brooks/Cole, Approach. Mediterranean Journal of Social
USA. Sciences, 5 (7):107.
5. Chatfield, C. & Yar, M. 1988a. Holt-Winters 21. Mushtaq, R. 2011. Augmented Dickey Fuller Test.
forecasting: some practical issues. The 22. Naudé, W. 2009. The financial crisis of 2008 and
Statistician:129-140. the developing countries. UNU-WIDER Discussion
6. Chatfield, C. & Yar, M. 1988b. Holt-Winters Paper No. 2009/01.
Forecasting: Some Practical Issues. Journal of the 23. Pîrvu, D. & Necşulescu, C. 2013. FUEL
Royal Statistical Society. Series D (The Statistician), CONSUMPTION: A MAJOR FACTOR INFLUENCING
37 (2):129-140. SALES OF NEW PERSONAL VEHICLES. EVIDENCE
7. Chifurira, R., Mudhombo, I., Chikobvu, M. & FROM ROMANIAN DATA. Studia Universitatis
Dubihlela, D. 2014. The Impact of Inflation on the Vasile Goldiş, Arad-Seria Ştiinţe Economice, (1):39-
Automobile Sales in South Africa. Mediterranean 52.
Journal of Social Sciences, 5 (7):200. 24. Sadowski, E.A. 2010. A Time Series Analysis:
8. Cipra, T. & Hanzák, T. 2008. Exponential smoothing Exploring the Link between Human Activity and
for irregular time series. Kybernetika, 44 (3):385- Blood Glucose Fluctuation.
399. 25. Schwarz, G. 1978. Estimating the dimension of a
9. Cipra, T., Trujillo, J. & Robio, A. 1995. Holt-Winters model. The annals of statistics, 6 (2):461-464.
method with missing observations. Management 26. Sean, P.M., Hill, K. & Swiecki, B. 2003. Economic
Science, 41 (1):174-178. Contribution of the Automotive Industry to the US
10. DA VEIGA, C.P., DA VEIGA, C.R.P., CATAPAN, A., Economy-An Update: A report for the Alliance of
TORTATO, U. & DA SILVA, W.V. 2014. Demand Automobile Manufacturers, Center for Automotive
Forecasting in Food Retail: A Comparison Between Research (CAR).
the Holt-Winters and ARIMA Models. WSEAS 27. Shahabuddin, S. 2009. Forecasting automobile
sales. Management Research News, 32 (7):670-682.

81
 Risk governance & control: financial markets & institutions / Volume 6, Issue 1, Winter 2016

28. Sivak, M. & Tsimhoni, O. 2008. Future demand for countries: a global value chain perspective. World
new cars in developing countries: going beyond Bank Policy Research Working Paper Series.
GDP and population size. 31. Williams, B.M. & Hoel, L.A. 2003. Modeling and
29. Stadnytska, T., Braun, S. & Werner, J. 2008. Model forecasting vehicular traffic flow as a seasonal
identification of integrated ARMA processes. ARIMA process: Theoretical basis and empirical
Multivariate Behavioral Research, 43 (1):1-28. results. Journal of transportation engineering, 129
30. Sturgeon, T. & Van Biesebroeck, J. 2010. Effects of (6):664-672.
the crisis on the automotive industry in developing 32. Yaffee, R. & McGee, M. 2000. Introduction to Time
Series Analysis and Forecasting with Applications
of SAS and SPSS. San Diego. Cal: Academic Press.

82