Journal of ICT, 18, No.

1 (January) 2019, pp: 1–18

How to cite this paper:

Tran, Q. T., Li, H., & Trinh, Q. K. (2019). Cellular network traffic prediction using
exponential smoothing Methods. Journal of Information and Communication Technology,
18 (1), 1-18.

CELLULAR NETWORK TRAFFIC PREDICTION

USING EXPONENTIAL SMOOTHING METHODS

Quang Thanh Tran, 1Li Hao & 2Quang Khai Trinh

1,2

Key Lab of Information Coding and Transmission, Southwest Jiaotong

1

University, China
2
Faculty of Electrical-Electronic Engineering, University of Transport
and Communications, Vietnam

thanhtq@utc.edu.vn; lhao@swjtu.edu.cn; khaitq@utc.edu.vn

ABSTRACT

Wireless traffic prediction plays an important role in network

planning and management, especially for real-time decision
making and short-term prediction. Systems require high accuracy,
low cost, and low computational complexity prediction methods.
Although exponential smoothing is an effective method, there is
a lack of use with cellular networks and research on data traffic.
The accuracy and suitability of this method need to be evaluated
using several types of traffic. Thus, this study introduces the
application of exponential smoothing as a method of adaptive
forecasting of cellular network traffic for cases of voice (in Erlang)
and data (in megabytes or gigabytes). Simple and Error, Trend,
Seasonal (ETS) methods are used for exponential smoothing. By
investigating the effect of their smoothing factors in describing
cellular network traffic, the accuracy of forecast using each
method is evaluated. This research comprises a comprehensive
analysis approach using multiple case study comparisons to
determine the best fit model. Different exponential smoothing
models are evaluated for various traffic types in different time
scales. The experiments are implemented on real data from a
commercial cellular network, which is divided into a training data
part for modeling and test data part for forecasting comparison.
Received: 30 January 2018 Accepted: 15 October 2018 Published: 11 December 2018

1
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

This study found that ETS framework is not suitable for hourly
voice traffic, but it provides nearly the same results with Holt–
Winter’s multiplicative seasonal (HWMS) in both cases of daily
voice and data traffic. HWMS is presumably encompassed by
ETC framework and shows good results in all cases of traffic.
Therefore, HWMS is recommended for cellular network traffic
prediction due to its simplicity and high accuracy.

Keywords: Cellular network traffic, exponential smoothing, Holt–Winter’s

multiplicative seasonal, wireless traffic prediction.

INTRODUCTION

Wireless traffic prediction is a key component of network planning, development,

and management. Accurate prediction will become even more necessary with
the development of 5th generation wireless systems (5G) that contain many
new service capabilities (5G PPP, 2015). The 5G system has a higher capacity
and higher density of mobile broadband users than the current 4G system.
It also supports device-to-device communications and massive machine-type
communications (NGMN Alliance, 2015). Consequently, people are living in
the age of social networks (Tyagi & Kumar, 2017) and the Internet-of-Things
(Matta, Pant, & Arora, 2017). Life becomes more convenient and intelligent
when everything can be connected via heterogeneous wireless networks
(Qiang, Li, & Altman, 2017). Along with these advanced technologies, Yusuf-
Asaju, Dahalin, and Ta’a (2018) also figured out the issues of mobile network
performance and proposed a framework for modeling mobile network quality
of experience using the big data analytics approach. And in fact, better network
operation and management are required to ensure a robust infrastructure that
includes the underlying network and supporting technologies, for example.
Analysis of wireless network traffic shows that the traffic series normally
contains seasonal components and can be modeled and forecasted by time
series analysis models (Tran, Ma, Li, Hao, & Trinh, 2015). Authors in these
papers proposed combining statistical procedures for modeling and forecasting
cellular network traffic, such as the autoregressive integrated moving average
(ARIMA) and generalized autoregressive conditional heteroskedasticity
(GARCH). They took advantage of the ARIMA model for capturing the
conditional mean of the traffic series and the GARCH model for dealing with
the conditional heteroskedasticity existing inside the traffic. They achieved
better forecast results compared with the individual models, but at the cost of
computational complexity. The results can be used for capacity planning and
overload warning issues that are important parts of network planning.
Exponential smoothing is a simple method of adaptive forecasting
in which the forecasts adjust based on past errors, unlike forecasts from
regression models that use fixed coefficients. Exponential smoothing

2
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

methods have been applied in several areas, such as palm oil real production
forecasting (Siregar, Butar-Butar, Rahmat, Andayani, & Fahmi, 2017), power
(Usaratniwart, Sirisukprasert, Hatti, & Hagiwara, 2017), revenue forecasting
(Rahman, Salma, Hossain, & Khan, 2016), and solar irradiance prediction
(Margaret & Jose, 2015), to name a few. These researchers all achieved good
results with this low-complexity and low-cost method. In terms of wireless
traffic prediction, Tikunov and Nishimura (2007) proposed the application of
Holt–Winter’s exponential smoothing, which is simple, low cost, does not
require a without highly skilled analyst, and operates nearly automatically
for GSM/GPRS network Erlang traffic prediction. The recorded data were
classified into three types, namely high, medium, and low intensity traffic
cells. The authors focused on cells with high and medium traffic intensity
for the purposes of overload warning and capacity planning. Although good
results were achieved, only voice traffic was considered. In the era of data,
there is a necessity for more comprehensive studies about using exponential
Base on the mentioned requirement, more exponential smoothing methods were investigated that
Base on theinmentioned
smoothing requirement,
cellular network more
traffic thatexponential
includes not smoothing
only voice methods werebut
(Erlang) investigated that
Base
also on
data the mentioned requirement, more exponential smoothing methods were investigated tha
onlyBase theon(megabytes
specific
the mentioned orrequirement,
gigabytes).
Holt–Winter’s more exponential
multiplicative smoothing
seasonal method
methods(HWMS), but also
were investigated thatdiffe
inclu
only the Base specific
on Holt–Winter’s
the mentioned multiplicativemore
requirement, seasonal method (HWMS),
exponential smoothing but also diffe
only the
onlyBasethe
specific
on
specific
Holt–Winter’s
the mentioned requirement,
Holt–Winter’s more exponential
multiplicative smoothing
seasonal method
methodmethods (HWMS), but
were investigated
(HWMS), but also
also diffe
that inclu
different
exponential
methods were smoothing methods.
investigated They
thatThey were notthenonly
multiplicative
included applied
seasonal
thetotospecific
forecast cellular network
Holt–Winter’s traffic that c
exponential smoothing methods. were then applied forecast cellular network traffic that ty
c
exponential
only
multiplicative smoothing
the specific methods.
Holt–Winter’s
seasonalmethods. They were then
multiplicative applied
seasonalto forecast
method cellular
(HWMS), network
but traffic
also that c
different
onlyexponential
voice (insmoothing
Erlang) butmethod
alsoTheydata(HWMS),
were
(in then
megabytes but also
applied to different
orforecast Intypes
cellular
gigabytes). network oftrafficthe
this study, thatsimple
consis
only voice
exponential (in
exponential Erlang)
smoothing but
smoothing also data
methods.
methods. They(in
They megabytes
were
were thenthen orto gigabytes).
applied
applied to forecast
forecast In network
cellular this study,
cellular trafficthethat simple
consist
only
only voice
voice (in(inErlang)
Erlang) but
but also
also data
data (in
(in megabytes
megabytes or orgigabytes).
gigabytes).In In
thisthis study,
study, the the
simple simple
exp
smoothingtraffic
network methods thatinclude
consistssingle,
of not double,
only Holt–Winter’s
voice (in Erlang) but alsoHolt–Winter’s
no seasonal, data (in additive
smoothing only methods
voice (in include single,
Erlang) but alsodouble,
data (in Holt–Winter’s
megabytes or gigabytes).
no In this
seasonal, study, the simple
Holt–Winter’s additiveexpo
smoothing
smoothingmethods
megabytes ormethods include
includeIn
gigabytes). single,
single, double,
thisdouble, Holt–Winter’s
study, Holt–Winter’s
the simple exponential
nono seasonal,
seasonal, smoothing
Holt–Winter’s
Holt–Winter’s additiveadditive
seaso
HWMS. The
smoothing methods
methods are introduced
include single, together
double, with
Holt–Winter’s an Error, Trend, Seasonaladditive(ETS)season fram
HWMS.
methods The
includemethods
single,are introduced
double, together
Holt–Winter’s withno an Error,
no Trend,
seasonal,
seasonal, Seasonal
Holt–Winter’s
Holt–Winter’s (ETS) fram
HWMS.
HWMS.The Themethods
methods areare introduced together
introduced together with
with ananError,
Error, Trend,
Trend, Seasonal
Seasonal (ETS) (ETS) fram
framewo
exponential
additive HWMS. smoothing
seasonal, The and methods
HWMS.
methods are
Theconsidered
are introduced methods
togetherinarethree cases
introduced
with an Error,of hourly
Trend,voice,
together with daily
Seasonal voice,framewor
an(ETS) and dai
exponential
Trend,smoothing
exponential smoothingmethods
methods are
are considered
considered inin threecases
three cases ofof hourly
hourly voice,
voice, dailydaily
voice,voice, and dai
and daily da
exponential
Error, smoothing
Seasonal methods
(ETS) are considered
framework. The inin three cases
exponential of hourly
smoothing voice, daily
methods voice, and dat
dai
types.
are exponential smoothing methods are considered three cases
considered in three cases of hourly voice, daily voice, and daily data traffic of hourly voice, daily voice, and daily
types.
types.
types.
types. types.
SIMPLE EXPONENTIAL SMOOTHING
SIMPLE EXPONENTIAL
SIMPLE EXPONENTIAL SMOOTHING
SMOOTHING
SIMPLE
SIMPLEEXPONENTIAL
EXPONENTIAL SMOOTHING SMOOTHING
Simple
Simple exponential
exponential smoothing
SIMPLE methods
EXPONENTIAL
smoothing methods include:
include:
SMOOTHING
Simple exponential smoothing methods include:
SimpleSimple exponential
exponential smoothing
smoothing methods
methods include:
include:
Simple exponential smoothing methods include:
- Single smoothing: one parameter
- - Single-Single smoothing: oneone parameter
Single smoothing: parameter 00<<<𝛼𝛼𝛼≤≤
00 1,
≤1,1,
smoothing: one parameter <𝛼≤ 1,
- Single smoothing:
- Single smoothing: one parameter one parameter 0 < 𝛼 ≤ 1,
- - Double-Double smoothing:
Doublesmoothing:
smoothing:one
oneone parameter
parameter
parameter 000<<<𝛼𝛼𝛼≤≤ ≤1,1,
1,
-- Double
Double smoothing:
smoothing: one one parameter
parameter 0 < 𝛼 ≤ 1,
- Double smoothing: one parameter 0 < 𝛼 ≤ 1,
- Holt-Winters
-Holt-Winters
Holt-WintersNo seasonal:
– No two
seasonal: parameters
two parameters000<<
- -- Holt-Winters ––No seasonal: two parameters <𝛼,𝛼,𝛼,
𝛽𝛽𝛽<<<
1,1,1,
Holt-Winters –– No
No seasonal:
seasonal: two
two parameters
parameters 0 < 𝛼, 𝛽 < 1,
- Holt-Winters – No seasonal: two parameters 0 < 𝛼, 𝛽 < 1,
- Holt-Winters – Additive seasonal: three parameters 0 < 𝛼, 𝛽, 𝛾 < 1, and
- - - Holt-Winters
Holt-Winters – Additive seasonal: three parameters 0 < 𝛼, 𝛽, 𝛾 < 1, and
Holt-Winters––AdditiveAdditiveseasonal:
seasonal:threethreeparameters
parameters 0 < 𝛼, 𝛽, 𝛾 < and 1, and
- Holt-Winters – Additive seasonal: three parameters 0 < 𝛼, 𝛽, 𝛾 < 1, and
- Holt-Winters – Additive seasonal: three parameters 0 < 𝛼, 𝛽, 𝛾 < 1, and
-Holt-Winters
Holt-Winters
- - Holt-Winters – Multiplicative
Multiplicative
– –Multiplicative seasonal:
seasonal:
seasonal: threeparameters
three
three parameters00<
parameters < 𝛼,
𝛼,𝛽,
𝛽, 𝛾𝛾 <
<11
- Holt-Winters – Multiplicative seasonal: three parameters 0 < 𝛼, 𝛽, 𝛾 < 1
- Holt-Winters – Multiplicative seasonal: three parameters 0 < 𝛼, 𝛽, 𝛾 < 1
- where
Holt-Winters
β,γand Multiplicative
–are the damping, seasonal: three parameters 0 < 𝛼, 𝛽, 𝛾 < 1
where
where α, α,
β, β,α,and
and are γthe
γ are thedamping,
damping, or or
or smoothing,
smoothing,
smoothing, factors.
factors.
factors.
where α, β, and γ are the damping, or smoothing, factors.
where α, β, and γ are the damping, or smoothing, factors.
where The and γ are
α, β,analysis of the damping,
cellular networkor 3smoothing, factors. for the HWMS method (Valakevic
traffic is appropriate
The analysis of cellular network traffic is appropriate for the HWMS method (Valakev
Brazenas,of2015),
The analysis whichnetwork
cellular is suitable for a series
traffic with a linear
is appropriate time
thetrend
fortime HWMS and multiplicative
method (Valse
TheBrazenas,
analysis2015), which isnetwork
of cellular suitable traffic
for a series with a linearfor
is appropriate thetrend
HWMS and multiplicative
method (Vals
The analysis
variation.of
If xcellular network traffic is appropriate for the
t is the input traffic series, then the smoothed series, 𝑥� , isHWMS
given by method (Val
 where
le smoothing: one
- α, β, and γ are
parameter
Holt-Winters 0the
– No < damping, or smoothing,
𝛼 ≤two1,parameters
seasonal: factors.
0 < 𝛼, 𝛽 < 1,
where α, β, and γ are the damping, or smoothing, factors.
where
where α, β, β, andand γγ are the
the damping, or smoothing, factors.
- α, Holt-Winters –are
Journal
Additive damping,
of ICT, 18,
seasonal: or smoothing,
No.parameters
three 1 (January) factors.
0 < 2019, pp:1,1–18
and
The
ble smoothing: analysis
one parameter of cellular0 < 𝛼network
≤ 1, traffic is appropriate
𝛼, 𝛽, 𝛾 <
for the HWMS method (Valakevi
The analysis of cellular network traffic is appropriate for the HWMS method (Valakevi
Brazenas,
The 2015), –which isnetwork
suitable for
threeaparameters
series with
0 < 𝛼,a𝛽, linear 1 timeHWMS trend and multiplicative se
The analysis of
of cellular
seasonal:traffic is
is appropriate
𝛾 <for
for the method (Valakevi
- Holt-Winters Multiplicative
analysis
Brazenas, 2015), cellular
which is network
suitable traffic
for a seriesappropriate
with a linear the HWMS
time trend and method (Valakevi
multiplicative se
The analysis of cellular network traffic is appropriate for the HWMS method
variation.
-Winters – Brazenas,
No seasonal:
Brazenas,
where α, β,If x t γis
2015), two
2015), which
and are thethe input
parameters
which damping, istraffic
suitable
is suitable
or 0series,
<
smoothing,for𝛼, then
a
𝛽
factors.<
for a series the
series
1, smoothed
with
with for a series,
linear
a linear time𝑥
time �𝑡 , is
trendgivenand
trendlinear bymultiplicative
and multiplicative s se
(Valakevicius
variation. If xt is&the Brazenas,
input traffic 2015), which
series, thenisthesuitable
smoothed aseries,
series 𝑥with
�𝑡 , is agiven by
variation.
time trend
variation. If x
and is the input
multiplicative
If xoftt iscellular
the input traffic
traffic series,
seasonal then the
variation. smoothed
If x is theseries,
input 𝑥traffic
� , is given
series, by
The analysis network trafficseries, then the
is appropriate smoothed
for the HWMS
t series,(Valakevicius
method 𝑥�𝑡𝑡 , is given & by
-Winters – Additive = (𝑎 +series,
seasonal: three (1)
suitableparameters
then the 𝑥�𝑡+𝑖smoothed 𝑏𝑖)𝑐 𝑡+𝑖 is given by0 < 𝛼, 𝛽, 𝛾 < 1, and
Brazenas,𝑥�𝑡+𝑖2015),
= (𝑎 which + is𝑏𝑖)𝑐 for a series with a linear time trend and multiplicative seasonal (1)
(1)
𝑡+𝑖
𝑥𝑥
variation. If xt =
��𝑡+𝑖 is
= (𝑎
the
(𝑎 +
input
+ 𝑏𝑖)𝑐
traffic𝑡+𝑖
𝑏𝑖)𝑐 series, then the smoothed series, 𝑥�𝑡 , is given by (1)
where a is𝑡+𝑖the permanent𝑡+𝑖 component (intercept), b is the trend, and (1) ct is the multiplicative s
-Winters – where
Multiplicative
a 𝑥�is seasonal: component
the permanent three parameters (intercept),0 < b𝛼,is𝛽,the 𝛾< 1 and
trend, (1) ct is the multiplicative s
factor.
where These three
𝑡+𝑖 = (𝑎 + 𝑏𝑖)𝑐𝑡+𝑖
coefficients are defined by the following recursions:
where a is the permanent component (intercept), b is the trend, and
where a is
a isthe
the permanent
permanent component
component (intercept),
(intercept), b bis isthe
the trend,
trend, andccct t is
and is the
the multiplicative se
t is the multiplicative s
factor. These three
multiplicative coefficients
seasonal factor. areThese
defined by the
three followingare
coefficients recursions:
defined by the
and γ are thefactor.
whereThese
a is the three
damping,
factor.
permanent
or
Theserecursions:
three coefficients
smoothing,
coefficients are
are defined
component (intercept),
factors. defined by by the
the following
b is the trend,
following recursions:
and ct is the multiplicative seasonal
recursions:
following
factor. These three
𝑎(𝑡) = 𝛼 (𝑡−𝑠)
𝑥𝑡
coefficients are defined by the following
+ (1 − 𝛼)�𝑎(𝑡 − 1)� + 𝑏(𝑡 − 1) recursions: (2)
𝑐𝑡 𝑥𝑡
𝑎(𝑡) = 𝛼 (𝑡−𝑠) 𝑥 + (1 − 𝛼)�𝑎(𝑡 − 1)� + 𝑏(𝑡 − 1) (2)
is of cellular network 𝑎(𝑡)
𝑎(𝑡) ==
𝑎(𝑡) = 𝛼𝛼𝑥𝑐𝑡 𝑡(𝑡−𝑠)
𝛼traffic 𝑐
𝑥+𝑡𝑡 (1 −
is ++ (1
appropriate
(1
𝛼)�𝑎(𝑡 −
− −𝛼)�𝑎(𝑡
𝛼)�𝑎(𝑡
1)� + 𝑏(𝑡 for
−
− 1)the
−1)�
1)� +
+ 𝑏(𝑡
HWMS
𝑏(𝑡 −− 1)
1) method (2)
(2) (Valakevicius & (2)
(2)
𝑐𝑡 (𝑡−𝑠)
𝑐𝑡𝑡 (𝑡−𝑠)
𝑏(𝑡) = 𝛽�𝑎(𝑡) − 𝑎(𝑡 − 1)� + (1 − 𝛽)𝑏(𝑡 − 1) (3)
015), which is suitable 𝑏(𝑡)
𝑏(𝑡) ==
for
𝛽�𝑎(𝑡)
𝛽�𝑎(𝑡)
a− 𝑎(𝑡 series
−− 𝑎(𝑡 with
1)� +− (11)� a+
− 𝛽)𝑏(𝑡
linear
(1
− 1)− time
𝛽)𝑏(𝑡trend − 1) and multiplicative(3) (3)
seasonal(3)
𝑏(𝑡) = 𝛽�𝑎(𝑡) − 𝑎(𝑡 − 1)� + (1 − 𝛽)𝑏(𝑡 − 1) (3)
(3)
xt is the input traffic 𝑏(𝑡) series,
= 𝛽�𝑎(𝑡) then
𝑥𝑡 −the 𝑎(𝑡smoothed
− 1)� + (1 series,
− 𝛽)𝑏(𝑡 𝑥
� , is
− given
1) by (4)
(4) (4)
(𝑡)==
𝑐𝑐𝑡𝑡(𝑡) 𝛾𝑡 𝑎(𝑡)
𝑥 + (1 − 𝛾)𝑐 (𝑡 − 𝑠) 𝑡
𝛾 𝑎(𝑡) 𝑥𝑡(1 − 𝛾)𝑐𝑡 (𝑡 − 𝑠) 𝑡
+
𝑐𝑡 (𝑡) = 𝛾 𝑎(𝑡) 𝑥𝑡 + (1 − 𝛾)𝑐𝑡 (𝑡 − 𝑠) (4)
where 𝑐𝑐𝑡 (𝑡)
(𝑡) = 𝛾𝛾 𝑎(𝑡)𝑥𝑡 + (1 − 𝛾)𝑐𝑡 (𝑡 − 𝑠)
are (1 the damping(𝑡 factors and s is the seasonal frequency. (4)
(4)
𝑖 = (𝑎 + 𝑏𝑖)𝑐 where0
=
<𝑡 𝛼, 𝛽, and 𝛾 𝑎(𝑡)
+ − 𝛾)𝑐 − 𝑠)
< 1 are the damping 𝑡factors and s is the seasonal frequency. (1)
where0
𝑡+𝑖 < 𝛼, 𝛽, and 𝛾 < 1 are the damping factors and s is the seasonal frequency.
where0
The
Theforecasts
< 𝛼, 𝛽,are
forecasts and
are 𝛾 < 1 are
computed
computed by: the damping factors and s is the seasonal frequency.
by:
where0 < 𝛼, 𝛽, and 𝛾𝛾 < 11 are the damping factors and ss is the seasonal
seasonal frequency. 3
the permanent component (intercept), the
where0 < 𝛼, 𝛽, and < are b isdamping
the trend, factors
andand ct is isthe themultiplicative frequency.seasonal
(5)
e three coefficients𝑥�𝑡+𝑖 are = defined
(𝑎(𝑇) + by𝑏(𝑇)𝑖)𝑐
the following
𝑇+𝑖−𝑠 recursions: (5)
where the seasonal factors are used from the last s estimates.
𝑥𝑡 where the seasonal factors are used from the last s estimates.
)= 𝛼 (𝑡−𝑠)
𝑐𝑡
+ (1 − 𝛼)�𝑎(𝑡 − 1)� + 𝑏(𝑡 − 1) (2)
ERROR TREND SEASONAL EXPONENTIAL SMOOTHING

) = 𝛽�𝑎(𝑡) − 𝑎(𝑡framework
This (1 − 𝛽)𝑏(𝑡
− 1)� + defines (3)
− 1) class of exponential smoothing methods
an extended
ETSfor
and offers a theoretical foundation EXPONENTIAL SMOOTHING
analysis of these models using state-space
𝑥𝑡 based likelihood calculations. Support for model selection and calculation of
𝑡) = 𝛾 𝑎(𝑡) This
+ (1 framework
− 𝛾)𝑐
forecast defines
𝑡 (𝑡 −
standard 𝑠) an
errors areextended classThe
also included. of exponential
standard exponential (4)
smoothingsmoothing
methods and offers a the
foundation for analysis
models discussed in theofprevious
these models using
section, suchstate-space
as HWMS,based likelihood calculations. Sup
are encompassed
by this ETS framework.
1 areselection
model
𝛼, 𝛽, and 𝛾 < Inthe
thedamping factors
and calculation
ETS exponential and s ismethod,
the standard
of forecast
smoothing seasonal
the timefrequency.
errors are also included. The standard expo
series may be decomposed
smoothing
into three models discussed
components, in the
namely the error
previous section,
(E) that such
is the as HWMS,
irregular are encompassed by th
unpredictable
component of the series, the trend (T) that characterizes the long-term movement
framework. 3
of the time series, and the season (S) that corresponds to a pattern with known
periodicity.
In the ETS exponential smoothing method, the time series may be decomposed into
components, namely the error (E) that is the irregular unpredictable component of the series, th
(T) that characterizes the long-term movement of the time series, and the season (S) that corre
to a pattern with known periodicity.
4
The ETS models can be described as state equations that are extended versions of those outl
 hod, the time components,
series may namely the error (E) that
be decomposed intois the irregular unpredictable component of the series,
components, components,
namely thenamely error (E) the that
erroris(E) thethat isthree
irregular theunpredictable
irregular unpredictable
componentcomponent of the
of the series, theseries,
trend
the error (E) components,
thatthat
(T) is the irregular
namelyunpredictable
characterizes the
the error
long-term(E) component
thatmovement ofof
is the irregularthethe series, the
unpredictable
time trend
series, component
and the season of(S)
thethat
series,
cor
e irregular unpredictable component
(T) that characterizes of the
the long-termseries, the trend
(T) that characterizes the long-term Journal
movementmovement
of ICT, 18,
of the timeofseries,
No. 1 (January) the time
2019,
andseries,
pp: 1–18 and the
the season (S)season (S) that cor
that corresponds
the long-term (T)movement
to that
aseries, of the
characterizes
pattern with time theseries,
long-term and movement
the season (S) of thethattime
corresponds
series, and the season (S) that cor
ent of the time towith
a pattern and with theknownseasonperiodicity.
known (S) that corresponds
periodicity.
to a pattern known periodicity.
wn periodicity. to a pattern with known periodicity.
The ETS models can be described as state equations that are extended versions
The ETS models can be described as state equations that are extended versions of those ou
The ETS The modelsofETS
those
can outlined
models
be describedcanbybeHyndman described
as state equations etasal.state
(2002) thatasare
equationsin extended
Equations
that are versions(6), (7), of
extended and (8), outlined
versions
those of thosebyou
be describedThe asrespectively:
Hyndman state
ETS equations
models
et al. can
(2002) that beasare
in extended
described
Equations as versions
state
(6), of those
equations
(7), and (8), outlined
that by
are extended
respectively: versions of those ou
equations that are extended versions of those outlined by
HyndmanHyndmanet al. (2002) et al.as (2002)
in Equations as in Equations
(6), (7), and (6),(8),(7),respectively:
and (8), respectively:
2) as in EquationsHyndman (6), (7), and(2002)
et𝑙𝑠al. = (8),𝛼𝑃(𝑥 respectively:
as in Equations
, 𝑒 ) + (1 (6),
− (7), and
𝛼)𝑄(𝑥 (8), respectively:
𝑡−1 ) (6)
7), and (8), respectively: 𝑙 = 𝛼𝑃(𝑥 )
𝑡−1
(1 , 𝑒
𝑡
) + (1 − 𝛼)𝑄(𝑥
) ) (6) (6) (6)
𝑙𝑠 = 𝛼𝑃(𝑥 , 𝑒 + − 𝛼)𝑄(𝑥
𝛼𝑃(𝑥𝑡−1 , 𝑒𝑡 ) + (1 − 𝛼)𝑄(𝑥 𝑠 𝑡−1
𝑏𝑙𝑠𝑡 𝑡−1=) 𝛼𝑃(𝑥 𝑡
𝛽𝑅(𝑥𝑡−1 𝑡−1 𝑡
𝑡−1, ,𝑒𝑒𝑡𝑡))+ +(1 𝑡−1
(1𝜙 − −𝛼)𝑄(𝑥
𝛽)𝜙1 𝑏𝑡−1
𝜙
𝑡−12
)(6) (7)
(6)
)𝑄(𝑥𝑡−1 ) 𝑏𝑡 = 𝛽𝑅(𝑥 𝑏𝑡 𝜙= , 𝛽𝑅(𝑥
𝑒 ) + (1 (6)
,−𝑒𝑡𝛽)𝜙 ) + 1(1 𝑏 −2 𝛽)𝜙 𝑏
𝑡−1
𝜙 2
(7) (7) (7)
𝛽𝑅(𝑥𝑡−1𝜙2 , 𝑒𝑡 ) + (1 − 𝛽)𝜙𝑏 𝑠1𝑡𝑡𝑏𝑡−1 𝑡−1
= 𝛽𝑅(𝑥
2 𝑡𝛾𝑇(𝑥𝑡−1𝑡−1
, 𝑒, 𝑒𝑡 )𝑡 )++(1 (1−−𝛽)𝜙
𝑡−1 1
𝑏𝑡−1(7)
𝛾)𝑠1𝑡−� 𝜙
𝑡−1
2
(8)
(7)
(1(7)
𝑡−1
)𝜙1 𝑏𝑡−1 𝑠𝑡 = 𝑠𝛾𝑇(𝑥 𝑡 =𝑡−1 , 𝑒𝛾𝑇(𝑥 ) +𝑡−1 ,−𝑒𝑡𝛾)𝑠 ) +𝑡−�(1 − 𝛾)𝑠𝑡−� (8) (8)
𝛾𝑇(𝑥𝑡−1 , 𝑒𝑡 ) + (1 − 𝛾)𝑠𝑠𝑡−� 𝑡 = 𝛾𝑇(𝑥𝑡−1 , 𝑒𝑡 ) + (1 − 𝛾)𝑠𝑡−� (8)
𝑡
(8) (8)
where l is a level term, b is(8)
𝛾)𝑠𝑡−�
a growth term, and s is a seasonal term. Variables P, R, an
where l iswherea level l isterm,
a level b isterm, a growth b is aterm, growth andterm, s is aand s is a term.
seasonal seasonal term. Variables
Variables P, R, andP,T R, arean
erm, b is a growthwhere
functions term,ofa and
l iswhere level
the isa level
sisterm,
prediction a seasonal
b is error a growth term.
and Variables
term, states;
lagged and is
P,s Q
R, aisanda Taare
seasonal
function term.of Variables
the lagged P, R, an
state;
m, and s is functions
a seasonalofterm. l Variables error
theR,prediction
term,
P, R, and b is a T are states; Q is a function of the lagged state; 𝜙
growth
and lagged term, and s is seasonal term.
functions of Variables
the prediction P, errorT are
and andfunctions
lagged states; of Q is a function
the lagged
prediction of lagged
the lagged state;
Q 𝜙1 is the 𝜙
diction error damping
and lagged
functions of states;
parameter Qforis linear
the prediction a function
error trend andof the
lagged
models; states;
𝜙 is Qerror
state;
the is 𝜙a1and
damping is theparameter
function states;
of the for
lagged state; 𝜙
multiplicati
ed states; Q damping
is ais function
a function
parameter ofofthe thefor lagged
lagged
linear state;trend𝜙𝜙1models;
state; is the
is thedamping
2
is parameter
the damping forparameter
linear trend
for multiplicati
damping parameter for linear trend models; is the 𝜙damping parameter for multiplicative trend
for linear trend models;
damping
models;
models; parameter
and isis the
𝜙𝑒2𝑡 ~𝑖𝑖𝑑(0, the fordamping
linear
damping
𝜎 2)
is parameter
trend
prediction
parameter models;
2 for𝜙multiplicative
error.
for multiplicative
2
is the damping
2 Figure 1 trendparameter
presents
trend the
models; modelfor of
and multiplicati
ETS exp
s; 𝜙2 is the models;
dampingand parameter for𝜎multiplicative
is prediction trend error.11 Figure 1 the
presents theof
models; and 𝑒𝑡 ~𝑖𝑖𝑑(0, ) isis prediction
𝑒𝜎𝑡 2~𝑖𝑖𝑑(0, 2)
prediction error.
error. Figure
Figure presents
presents the model
model ofmodel
ETS of ETS exp
ETS exponential
(0, 𝜎 ) is prediction
2
models; error.𝑒𝑡smoothing
and Figure
~𝑖𝑖𝑑(0,1𝜎 presents ) is prediction the model of Figure
error. ETS exponential
1 presents the model of ETS exp
error. Figuresmoothing presentsmethod.
2
1 exponential the model ofmethod. ETS exponential
smoothingsmoothing
method. method.
smoothing method.

4 4
4
Figure 1. Model of ETS Exponential Smoothing Method.

Model Estimation

Model Specification

In this section, the type of ETS model used for smoothing is specified. There
are a total of 30 possible ETS models based on the various combinations of
the three components, E, T, and S, as defined in Equations (9), (10) and (11),
respectively:

5
 Journal ofEICT,
{A,18,
M}No. 1 (January) 2019, pp: 1–18
(9)
T {N, A, M, AD, MD}
E {A, M} (9)
(10)
S {N, A, M}
T {N, A, M, AD, MD} (10)
(11)
where N = none, A = additive, M = multiplicative,(11)
S {N, A, M} AD = additive dampened
ened, and MDwhere
= multiplicative
N = none, A dampened.
= additive, M = multiplicative, AD = additive dampened,
and MD = multiplicative dampened.
Estimation Control
Estimation Control
According to the chosen ETS model specification, the corresponding u
According to the
The Gaussian logchosen ETS for
likelihood model
ETSspecification,
specificationsthecancorresponding unknown
be written in terms of the prediction
ng unknown parameters 𝜃 =
parameters (𝛼, 𝛽, 𝛾, 𝜙), and
and the
theinitial
initialstates
states x
x00 may
may be
be estimated
estimated using
using either the maxim
in Equation (12):
either the maximum likelihood or average mean square error minimization
aximum likelihood
The
(AMSE) log mean
or average
Gaussian
methods.
square
likelihood forerror
ETSminimization
specifications(AMSE) methods.
can be written in terms of the prediction e
The
in Gaussian
Equation log likelihood for ETS specifications
1 can 𝑒be 2 written in terms of
𝑙𝑜𝑔(12):
2
𝐿(𝜃, 𝑥0 , 𝜎 2 |𝑦) = − 𝑙𝑜𝑔(2𝜋𝜎 2 ) − ∑�𝑡−1 𝑡2 . (12)
the prediction errors, as in Equation (12):
𝑇 2 𝜎

5 2 1 � 𝑒𝑡 2
The parameters
𝑙𝑜𝑔 𝐿(𝜃, 𝑥and
0 , 𝜎initial
2
|𝑦) =states are achieved
− 𝑙𝑜𝑔(2𝜋𝜎 2)
− 2by maximizing
∑𝑡−1
𝜎2
. the (12) (
likelihood in Equation
(12)
𝑇
respect
The to 𝜃, 𝑥0and
parameters , and usingare
𝜎 2 states
initial theachieved
Broyden, Fletcher, Goldfarb,
by maximizing and Shanno
the likelihood in (BFGS) a
The parameterswith and initial states are achieved by maximizing the likelihood in Equation (
(Fletcher,(12)
Equation 1987). respect to and using the Broyden, Fletcher, Goldfarb, and
respect
Shanno to(BFGS) and 𝜎 2 using
𝜃, 𝑥0 ,algorithm the 1987).
(Fletcher, Broyden, Fletcher, Goldfarb, and Shanno (BFGS) a
The average mean square error (AMSE) of the h-step forecasts is expressed
The average
(Fletcher, mean square error (AMSE) of the h-step forecasts is expressed as in Equation (1
1987).
as in Equation (13):

The average
𝐴𝑀𝑆𝐸mean
= square
∑�=1 �error
1 ℎ 1 ℎ
(AMSE)
∑�=1 2
𝑒𝑡+�|𝑡 �of. the h-step forecasts is expressed
(13) as in Equation
(13) (13
𝑇 ℎ

The parameters
1 ℎ and 1initial states that minimize the AMSE using BFGS
The parameters
𝐴𝑀𝑆𝐸 =and initial
� states
∑ ℎ
that
are then achieved.𝑇 �=1 ℎ �=1 𝑡+�|𝑡 �.
∑ 𝑒 2
minimize (13)
the AMSE using BFGS are then achieved.

Model Selection
(2) Model Selection
The parameters and initial states that minimize the AMSE using BFGS are then achieved.
In this step, the model can be selected based on either the comparison of a
In Model
(2) this step, theinformation
Selection
likelihood-based model cancriterion
be selected
acrossbased on either
the models the which
to decide comparison
one of a likeliho
information
more closely criterion across
fits the data theforecast
or the models error,
to decide which
i.e., an one more AMSE
out-of-sample closely to
fits the data or the
In this
decide step,
which the
one model
has can
a more be selected
accurate forecastbased on
result. either the comparison
error, i.e., an out-of-sample AMSE to decide which one has a more accurate forecast of a likelihoo
result.
The likelihood-based information criteria include Akaike information criterion
information
(AIC), Schwarz criterion across the
information models(BIC),
criterion to decide
or thewhich one more closely
Hannan-Quinn fits the data or the
criterion
The likelihood-based
error,
(HQ), i.e., an out-of-sample
as expressed information
AMSE
in Equations tocriteria
(14), decide
(15), andinclude
which
(16),oneAkaike information
has a more
respectively: accurate criterion (AIC),
forecast result.
information criterion (BIC), or the Hannan-Quinn criterion (HQ), as expressed in Equati
The
(15),likelihood-based information criteria include Akaike information criterion (AIC),
and (16), respectively:
information criterion (BIC), or the Hannan-Quinn criterion (HQ), as expressed in (14)
Equatio
𝐴𝐼𝐶 = −2 𝑙𝑜𝑔 𝐿�𝜃�, 𝑥�0 � + 2𝑝
6
(15), and (16),
𝐵𝐼𝐶 respectively:
= −2 𝑙𝑜𝑔 𝐿�𝜃�, 𝑥�0 � + 𝑙𝑜𝑔(𝑇)𝑝 (15)
𝐴𝐼𝐶
𝐻𝑄 == −2 𝑙𝑜𝑔 −2𝐿�𝜃�, 𝑥�𝐿�𝜃
𝑙𝑜𝑔 �
0� +, 𝑥�2 + 2𝑝 𝑇)𝑝
0 �𝑙𝑜𝑔(𝑙𝑜𝑔 (14)
(16)
 he Hannan-Quinn The likelihood-based
The likelihood-based
criterion (HQ), as information
expressed criteria
information criteria
in Equations include
include (14), Akaike information
Akaike information criterion criterion (AIC), (AIC),
information criterionand (BIC), orstates the Hannan-Quinn criterion (HQ), as expressed inthenEqua
The The parameters
likelihood-based initial
information criteria that minimize
include the
Akaike AMSE
information using BFGS
criterion are (AIC), a
nimize the AMSE information
informationusing criterion
BFGS
criterion are(BIC),
Journal
then
(BIC), or
ofThe
ICT, the
achieved.
or the
18, Hannan-Quinn
likelihood-based
No.
Hannan-Quinn
1 (January) 2019, criterion
informationpp: 1–18(HQ),
criterion criteria
(HQ), as
as expressed
include
expressed in
Akaike
in Equatio
inform
Equatio
The and
(15), likelihood-based
(16), respectively: information criteria include Akaike information criterion (AIC), S
information criterion (BIC), information or the Hannan-Quinn criterion (BIC), criterionor the(HQ), as expressed
Hannan-Quinn in Equatio
criterion (HQ)
𝑔e𝐿�𝜃 information
�, 𝑥�0 � + 2𝑝(15), and
criterion (16),(AIC), respectively:
Schwarz (14)
(15), (2)
andModel
information(16), 𝐴𝐼𝐶 Selection
respectively:
criterion
= (BIC), −2 or𝑙𝑜𝑔 the 𝐿�𝜃 Hannan-Quinn
� , 𝑥
� � + 2𝑝 criterion (HQ), as expressed in Equation
(14)
(15), and (16), respectively: (15), and �(16), respectively: 0
on (HQ), as expressed in Equations ��, ,(15) (14)
= (14),
�𝜃�, 𝑥 �0 � + 𝑙𝑜𝑔(𝑇)𝑝 𝐴𝐼𝐶 =respectively: −2𝑙𝑜𝑔 𝑙𝑜𝑔𝐿�𝜃 𝐿�𝜃
(15), and 𝐴𝐼𝐶 (16),
𝐵𝐼𝐶 = −2
−2 𝑙𝑜𝑔 𝐿�𝜃 𝑥,�𝑥𝑥�0�00���++ +𝑙𝑜𝑔(𝑇)𝑝
2𝑝
2𝑝 (14) (14) (15)
𝐴𝐼𝐶 = −2 𝑙𝑜𝑔 𝐿�𝜃 � , �𝑥�0+� 𝑙𝑜𝑔(𝑇)𝑝+=2𝑝 (14)
𝑥�0 � + 2 𝑙𝑜𝑔(𝑙𝑜𝑔 𝑇)𝑝 In this 𝐵𝐼𝐶
𝐵𝐼𝐶 step,=
== =the−2
−2
−2 model𝑙𝑜𝑔
𝑙𝑜𝑔 𝐿�𝜃
𝐿�𝜃
�
�can
�, ,𝑥�𝑥��(16)
, 𝑥� 𝐴𝐼𝐶
𝜃��be ,+ selected
𝑥�20𝑙𝑜𝑔(𝑇)𝑝
𝜃� , 𝑥� 0 � + 2𝑝
based on either the comparison
−2 𝑙𝑜𝑔 𝐿 � (15)
(14)
(15) o
(16)
𝐻𝑄 𝑙𝑜𝑔 𝐿�𝜃 0 + �𝑙𝑜𝑔(𝑙𝑜𝑔
+ 2𝑝 𝑇)𝑝
𝐴𝐼𝐶 −2 𝑙𝑜𝑔
�, 𝐿𝑥�000�𝐵𝐼𝐶 (15)
ed based on either 𝐵𝐼𝐶 𝐻𝑄the
(14) comparison
== −2 −2𝑙𝑜𝑔 𝑙𝑜𝑔𝐿�𝜃of𝐿�𝜃
� , a
𝑥
� likelihood-based
� + � 2+𝑙𝑜𝑔(𝑙𝑜𝑔𝑙𝑜𝑔(𝑇)𝑝
= −2
𝑇)𝑝 𝑙𝑜𝑔 𝐿 � �, 𝑥�0 � + 𝑙𝑜𝑔
𝜃 (15) (𝑇 ) 𝑝 (16)
information𝐻𝑄 𝐵𝐼𝐶= criterion
=−2 𝑙𝑜𝑔 −2across
𝐿�𝜃𝑙𝑜𝑔
�, 𝑥�00�the
�, 𝑥 𝐿��0𝜃 ��,+ 𝑥�02models� 𝑙𝑜𝑔(𝑙𝑜𝑔
+ 𝑙𝑜𝑔(𝑇 to)𝑇)𝑝
𝑝decide which one more closely
�, 𝑥� 0 � + 2 𝑙𝑜𝑔(𝑙𝑜𝑔 𝑇)𝑝
(15)
(16)fits th
values
to decide and which
p is the onenumber 𝐻𝑄
more
(15)𝐻𝑄of =
closely
=
−2
parameters −2
𝑙𝑜𝑔
fits𝑙𝑜𝑔
𝐿�𝜃
inthe𝐿�
𝜃� plus
𝜃data
� , 𝑥 �
+
the
�
𝐻𝑄
or+
2 𝑙𝑜𝑔(𝑙𝑜𝑔
number
2the
𝑙𝑜𝑔
= −2of
forecast
( 𝑙𝑜𝑔
𝑇)𝑝
𝑇
𝑙𝑜𝑔
the
) 𝑝
𝐿 � 𝜃 (16)
where
error, �𝜃�,i.e.,
𝑥�0 � an areout-of-sample
the maximized AMSE values
0 and p is the which
to decide numberone of parameters a morein�accurate
has (16) 𝜃� plus (16)theforec
num
chosen model where
ishas
the
�
�𝜃�aone
�𝜃 , 𝑥� 0�with
, 𝑥�0(16) � are the maximized
the maximized
minimum values
value and
ofand AIC p is
p is(BIC,the number
or of parameters in 𝜃 plus the numb
cide which one
where
estimated �, 𝑥�more
areaccurate
initial the
states inforecast
0 . The
values result.
�𝜃chosen model theisnumber the onevaluesof parameters in 𝜃� plus the numb
where �𝜃 � are the maximized 𝑥�where values � , 𝑥� 0and � are p the
is the maximized
number ofwith the
and minimum
parameters p isin plusvalue
the𝜃�� number theofofnumbAI
para
where initial
estimated
where � �
0 � are
𝜃 , 𝑥� 0 states
are the
the inmaximized.
maximized
𝑥� The values
chosenvalues and
model
and p is
p the
is
is the
the number
one
number of
with parameters
ofthe minimum
parameters in 𝜃 plus
invalue the of numbe
AIC
HQ).The initial
estimated likelihood-based
states in 𝑥�00. The information
estimated
chosen model
initial
criteriais theinclude
statesisin inthe
one withAkaike
. The chosen
the minimum information
model is is the
value ofcriter
the one of with
AIC
the
r of parameters estimated plusinitial
inestimated
𝜃� plus the states
number
theinitial
number inthe
of
of 𝑥�0𝑥.�estimated
the The chosen initial model states 𝑥� 0.one
The with
chosenthe minimum
model value AIC
eria includeHQ). Akaike
HQ). information states criterion
in . The (AIC),
chosen modelSchwarz is the one with the minimum value of AIC (
information criterion (BIC), or the Hannan-Quinn criterion (HQ), as expresse
0
one with the minimum value of AIC (BIC, or HQ).
nwith
in Equation HQ).
the minimum (17):
HQ).value of AIC (BIC, or HQ).
nnan-Quinn criterion The out-of-sample
(HQ), as AMSE expressed is given in in EquationsEquation (17): (14),
The (15),
The and (16),AMSE
out-of-sample
out-of-sample respectively:
AMSE is given
is given in Equation
in Equation (17): (17):
The out-of-sample AMSE is given in Equation AMSE
The out-of-sample (17): is given in Equation (17):
The out-of-sample
The out-of-sample AMSE1AMSE is given in Equation (17):
𝑇+𝑇 ∗ is given ℎ in𝑙𝑜𝑔 Equation
(17) �, 𝑥(17):
2
=1 𝑒𝑡+�|𝑡 �. 𝐴𝐼𝐶
𝐴𝑀𝑆𝐸 =1 ∗ 𝑡=𝑇+1 ∑= �
1 −2
∑ 2 𝐿�𝜃
𝑒𝑡+�|𝑡 �. � � + 2𝑝 (17)
0 (17)
1 ∑𝑇+𝑇 ∗ �1 ∑ℎ (14)
1 ℎ ℎ �=1
𝑥�0 � + 2𝑝 (17)
𝑇 𝑇+𝑇 ∗ 2
𝐴𝑀𝑆𝐸
𝐴𝑀𝑆𝐸 = 𝑇1∗ 1𝑇+𝑇 =
𝐵𝐼𝐶 ∑ = ∗ ∗�1
−2
∑ 𝑙𝑜𝑔 𝑒 2
𝑒𝑡+�|𝑡 𝐿�𝜃 �
� ,. 𝑥
�
. 0∑ 1 � + 𝑙𝑜𝑔(𝑇)𝑝
∗ 1 (17)
2 2 �=
𝑡=𝑇+1 �=1 𝑡+�|𝑡 𝑇+𝑇 ℎ 2
∗ ℎ 1 �=1 𝐴𝑀𝑆𝐸 � ∑ 𝑒 �.
(17)
𝑡=𝑇+1
+ 𝑙𝑜𝑔(𝑇)𝑝 Smoothing/Forecasting 𝐴𝑀𝑆𝐸 𝐴𝑀𝑆𝐸 = 𝑇=∗ ∑∗𝑡=𝑇+1 ∑𝑇+𝑇
𝑡=𝑇+1 �ℎℎ�∑ℎ�=1 ∑ℎ𝑘=1 (15)
𝑒𝑡+�|𝑡
𝑒�𝑡+𝑘|𝑡 �.�. 𝑇∗ 𝑡=𝑇+1 ℎ 𝑘=1 𝑡+𝑘|𝑡 (17)
(3) Smoothing/Forecasting
(17)
𝐻𝑄 = −2 𝑙𝑜𝑔 𝐿�𝜃 , 𝑥�0 � + 2 𝑙𝑜𝑔(𝑙𝑜𝑔 𝑇)𝑝
𝑇 𝑇 ℎ
2 𝑙𝑜𝑔(𝑙𝑜𝑔 𝑇)𝑝 (3) Smoothing/Forecasting
(3) Using
Smoothing/Forecasting (16)
the chosen ETS (3) model Smoothing/Forecasting
with estimated parameters, the in-sample
with estimated (3)parameters,
Smoothing/Forecasting
(3) Smoothing/Forecasting
the in-sample smoothed series can be
Using
smoothed the chosen series ETS
can be model
achieved with byestimatedthe one-step-ahead parameters,forecast the in-sample function, smoothed �seri
Using
wherethe �𝜃�, 𝑥�0 �ETS
chosen
are the model
maximized
with estimated
values and
parameters,
p is the the number in-sample
of parameters
smoothed
inseries
𝜃p
orecast function, Using ℎ(𝑥 the chosen
, 𝜃). Based
Based ETS on
on the
model
the smoothing
Using with
smoothing the estimated
model model
chosen and and
ETS the
model
parameters,
the estimated with
the estimated
in-sample
parameters, as parameters,
smoothed
well the
seriesin
and p is the Using number
achieved
Using theof
𝑡−1
the parameters
by
chosen the one-step-ahead
chosen ETS ETS in
model 𝜃 plus
�
model with theestimated
forecast
with number function,
estimated of the
parameters,
ℎ(𝑥𝑡−1 , 𝜃).
parameters, theBased
the in-sample
in-sample on thesmoothed
smoothing
smoothed mod
series
series
asestimated
the
achieved by in-sample
by the initial data,
the one-step-ahead states
one-step-ahead the in
out-of-sample
𝑥 �
forecast . The chosen
forecasted
function, model valuesis the
can one
be with
obtained
Basedfunction,
on the the by
the smoothing minimum
smoothing mode va
be achieved
0 by the one-step-ahead
ℎ(𝑥𝑡−1,, 𝜃). forecast
𝜃). ℎ (𝑥 𝑡−1 , 𝜃).mode Based
s, the in-sample achieved
smoothed series can forecast function, ℎ(𝑥 Based on
n model is the one
achieved
achieved with
by
implementing by
thethethe minimum
one-step-ahead
one-step-ahead
the dynamic valueforecast
forecast of of
forecast AICfunction,
the (BIC,
function,
series.ℎ(𝑥 ℎor(6 𝑥𝑡−1 , 𝜃). Based
𝑡−1, 𝜃). Based on on the
the smoothing
smoothing model
model
𝜃). Based on the smoothing
HQ). model and the estimated parameters, as well 𝑡−1
as the in-sample data, the out-of-
estimated parameters, as well as the in-sample data, the out-of-sample forecasted values
out-of-sample forecasted obtained by implementing the dynamic forecast of the series.
byvalues
obtained EXPERIMENTAL can be the
implementing
The out-of-sample AMSEdynamic is given
METHODOLOGY forecast of the series.
in EquationAND (17):RESULT
s.uation (17):
EXPERIMENTAL METHODOLOGY AND
The real traffic from EXPERIMENTALcommercial cellular METHODOLOGY networks wasAND RESULTS
collected and
1 𝑇+𝑇 ∗ 1 ℎ
GY AND RESULTS included voice 𝐴𝑀𝑆𝐸 in = Erlang ∑ and data ∑
in Mbps. 2 The first trace is 31-day hourly
.
𝑇 ∗ The real ℎtraffic
𝑡=𝑇+1 � �=1from 𝑒 𝑡+�|𝑡 �
commercial cellular networks was collected
The real traffic from commercial during the(17) cellular networks periodwas collected andofincluded
MSC_ voice in Erla
�|𝑡 �. voice traffic collected
data in Mbps.
New Year
The first trace
from
is 31-day
MSCID
hourly voice traffic collected
ollected and included data in
Hanoi1. voiceMbps.
The in traffic
Erlang
The first and
data trace rangedis 31-day fromhourly 00:00voice December traffic 9th, collected
2015during the New Year perio
until 23:00
collected during the (3)
New
January Smoothing/Forecasting
Year8th, period
2016 from and included MSCID of 744 MSC_Hanoi1.
observations. The traffic data trace
ranged is from
the 00:00 Decem
MSCID of MSC_Hanoi1. The traffic data ranged from The 00:00second
December 9th, 2015 until 23:00 J
0 December 9th, daily 20152016 voice
untiland trafficJanuary
23:00 and included 8th, 2016 236and 3G included
traffic values 744 observations.
that ranged from The second
March trace is the da
8th, included 744 observations. The second trace is the daily voice traffic and includ
Usingand
9th
is the daily voicedistrict
until
traffic
theincluded
Octoberchosen236ETS
30th, 2014, modelvalues
3G traffic
collected with that estimated
from 110
ranged from parameters,
siteBs
March 9ththe
in the
untilin-sample
CauGiay
October 30th, smo 2
3G traffic values
timated parameters, the in-sample smoothed series can be
area in that
Hanoi. ranged
The from
third March
trace is 9th
the until
3G data Octobercollected30th, 2014,
from 600collected
cells from 110 s
r 30th, 2014, collected achieved
located
the CauGiayfrom within by
110 the
the one-step-ahead
siteBs
district in the
HoanKiem
area
CauGiay
in Hanoi. district The in
forecast
district
third Hanoi. function,
area
trace The
in Hanoi. ℎ(𝑥
is the443-value
The
3G data𝑡−1 ,third
𝜃).
dailyBased
trace
collected data isonthethe 3Gsmoo
from 600 cells
data
function, ℎ(𝑥𝑡−1 , 𝜃).series
traffic Basedranged on thefrom smoothing
within
September the model HoanKiem12th,and theto in
district
2014 Hanoi. The
November 443-value
28th, 2015. daily data tra
3G data collected fromthe
within 600HoanKiem
cells located district in Hanoi. The 443-value daily data traffic series ranged from Sep
Exponential smoothing methods
12th, 2014 to November 28th, 2015. were applied to forecast these three
y data traffic series12th, ranged 2014 fromto September28th,
November 2015.
traces that are distinguished from each other by 6traffic type, namely voice
vs. data and hourly vs. daily. The simple exponential smoothing methods
Exponential smoothing methods were applied to forecast these three
Exponential
included smoothing
single, double, methods were applied and
and Holt-Winters to forecast
were used thesetothree traces that
implement theare distinguishe
ese three traces that areother
forecasts distinguished
forbythetraffic
threefrom each
trafficnamely other
traces. voice by
Then, vs. traffic
the data type,
achieved namely voice vs. data
theand hourly v
each type, and models
hourly vs. thatdaily.
gave The simple expo
hourly vs. daily. smoothing
The simplemethods exponential smoothing methods included single, double , and Holt-Winters a
included single, double , and Holt-Winters and were used to implem
Winters and wereforecasts used to for implement forecasts for 7the three traffic traces. Then, the achieved models that
the three the traffic traces. Then, the achieved models that gave the best results were
dels that gave thetobest results were chosen to make comparisons with the best choices by the ETS exponential
make comparisons with the best choices by the ETS exponential smoothing framework in t
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

best results were chosen to make comparisons with the best choices by the
ETS exponential smoothing framework in terms of estimation information,
in-sample forecast, and out-of-sample forecast.

Hourly Voice Traffic

The simple exponential smoothing methods were applied to hourly voice

traffic. The estimation outputs are shown in Table 1, where HWMS presents
the best results in terms of the 426.5057 RMSE. Thus, the HWMS was chosen
as the best forecasting method in the case of hourly voice traffic. In contrast,
the ETS framework was also applied to this hourly voice traffic, and the
estimation outputs are listed in Table 2. The numerical results indicate that this
framework is not suitable for the hourly data series when both AIC-based and
AMSE-based models yield abnormal RMSE values. To confirm this judgment,
we further implement the in-sample and out-of-sample forecast tests for these
two models and compare the results with those obtained using HWMS. The
results are shown in Figures 2 and 3 which display that the HWMS out-of-
sample forecasted data is very close to the original data, so we can figure out
that HWMS outperforms the two ETS models in modeling and forecasting
hourly cellular voice traffic.

Table 1

Hourly Voice Traffic Estimation Outputs of Simple Exponential Smoothing

Methods.

Single Double HWNS HWAS HWMS

Exponential Smoothing Methods

Alpha 0.9990 0.9990 1.0000 1.0000 1.0000

Parameters Beta 1.0000 0.0000 0.0000

Gamma 0.0000 0.0000

Sum of Squared Residuals 5.36E+09 4.15E+09 4.05E+09 2.36E+08 1.35E+08

Root Mean Squared Error 2684.505 2363.002 2334.268 563.0144 426.5057

Mean 2797.097 2793.238 2793.240 9059.919 10147.16

End of Period
Levels
Trend -3854.95 -3850.91 1.419014 1.419014

8
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

Table 2

Hourly Voice Traffic Estimation Outputs of AIC and AMSE Based ETS Chosen
Models.

AIC Based AMSE Based

Chosen Models

E E{M} E{A}
T T{MD} T{MD}
S S{N} S{M}
Parameters

Alpha 1.000000 1.000000

Beta 0.986463 0.000000
Gamma 0.000000
Phi 1.000000 1.000000
Estimation information

AIC 16065.87 16665.64

SSR 95.84155 5.25E+09
RMSE 0.358914 2656.054
AMSE 3.21E+40 21744630

2. In-sample
FigureFigure forecast by
2. In-sample HWMSby
forecast and ETS models
HWMS in case
and ETS of hourly
models voiceoftraffic
in case hourly(2-day view
window)
voice traffic (2-day view window)

9
Figure 2. In-sample forecast by HWMS
Journal andNo.
of ICT, 18, ETS models in2019,
1 (January) case of
pp:hourly
1–18 voice traffic (2-day view
window)

FigureFigure 3. Out-of-sample
3. Out-of-sample forecast byforecast
HWMS andbyETS
HWMS
modelsand
for ETS
1 day models
in case offor 1 day
hourly in traffic.
voice
case of hourly voice traffic.
(2) Daily Voice Traffic

The same experimental procedure was implemented for the case of daily voice traffic. Table 3
Daily Voice Traffic
illustrates the estimation outputs of the simple exponential smoothing methods. The HWMS again
presents
The the bestexperimental
same results among procedure
other methods
wasin implemented
term of RMSE,for
which
the is 531.6792.
case Thus,
of daily HWMS
voice
was traffic.
chosen asTable
the best
3 forecasting
illustratesmethod for the case outputs
the estimation of daily voice traffic.
of the The ETS
simple framework was
exponential
smoothing methods. The HWMS again presents the best results among other
methods in term of RMSE, which is 531.6792. Thus, HWMS was chosen 8
as the best forecasting method for the case of daily voice traffic. The ETS
framework was also applied to the daily voice traffic, and the estimation
outputs are illustrated in Table 4. Based on the results, the AIC-based ETS
model, i.e. {A, N, M}, should be chosen due to the better RMSE.

Table 3

Daily Voice Traffic Estimation Outputs of Simple Exponential Smoothing

Methods.
Exponential Smoothing Methods Single Double HWNS HWAS HWMS
Alpha 0.0010 0.0070 0.1100 0.8200 0.8400
Parameters Beta 0.0000 0.0000 0.0000
Gamma 0.0000 0.0000
Sum of Squared Residuals 3.84E+08 3.89E+08 4.33E+08 73565891 66713129

(continued)

10
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

Exponential Smoothing Methods Single Double HWNS HWAS HWMS

Root Mean Squared Error 1275.546 1283.683 1354.245 558.3188 531.6792

End of Period Levels Mean 5259.193 5440.615 5707.966 5691.035 5662.484
Levels Trend 2.008049 9.387678 1.317833 1.317833

Table 4

Daily Voice Traffic Estimation Output of AIC and AMSE Based ETS Chosen
Models.

AIC Based AMSE Based

Chosen Models

E E{A} E{A}
T T{N} T{A}
S S{M} S{M}
Parameters
Alpha 0.838877 0.000000
Beta 0.000000
Gamma 0.000000 0.000000
Phi
Estimation information
AIC 4269.541 4363.242
SSR 66652447 97473742
RMSE 531.4373 642.6697
AMSE 484775.5 412858.1

The comparison between the HWMS and ETS models was done next for the
case of daily voice traffic. It can be seen from Table 5 that the achieved ETS
model shows the same parameters as HWMS. In addition, there appears to be
no difference between the HWMS and ETS{A,N,M} in terms of in-sample
forecasts in Figures 4 and 5, and the out-of-sample forecast in Figure 6.

11
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

Table 5

Comparison of HWMS and ETS Methods in Case of Daily Voice Traffic.

Simple ETS
Exponential Smoothing Methods
HWMS {A,N,M}

Alpha 0.838877 0.838877

Beta
Parameters
Gamma 0.000000 0.000000
Phi

Akaike Information Criterion 4269.541

Sum of Squared Residuals 66713129 66652447
Root Mean Squared Error 531.6792 531.4373
Average Mean Squared Error 484775.5

Figure4. In-sample forecast by HWMS and ETS in case of daily voice

traffic.

12
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18
Figure4. In-sample forecast by HWMS and ETS in case of daily voice traffic.

5. In-sample
FigureFigure forecast byforecast
5. In-sample HWMS andbyETS for 3 weeks
HWMS in casefor
and ETS of 3
daily voicein
weeks traffic.
case of
daily voice traffic.

10

6. Out-of-sample
FigureFigure forecast byforecast
6. Out-of-sample HWMS andbyETS for 3 weeks
HWMS in case
and ETS forof3daily
weeksvoice
intraffic.
case
of daily voice traffic.
(3) Daily Data Traffic

The HWMS again is the best choice among other the simple exponential smoothing methods in the
case of daily data traffic due to the best RMSE 13
of 48.03080, as presented in Table 6. Besides, AIC-
based ETS {A,N,M} was also the chosen model with the lower RMSE of 48.03521, as shown in
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

Daily Data Traffic

The HWMS again is the best choice among other the simple exponential
smoothing methods in the case of daily data traffic due to the best RMSE
of 48.03080, as presented in Table 6. Besides, AIC-based ETS {A,N,M}
was also the chosen model with the lower RMSE of 48.03521, as shown
in Table 7. The same results were achieved in the case of daily data
traffic in which the HWMS and ETS{A,N,M} have almost the same
parameters, as in Table 8. The same is true for the same in-sample forecasts
in Figures 7 and 8 and the same out-of-sample forecast in Figure 9.

Table 6
Daily Data Traffic Estimation Outputs of Simple Exponential Smoothing
Methods.

Exponential Smoothing Single Double HWNS HWAS HWMS

Methods
Alpha 0.8080 0.1140 0.3200 0.6600 0.6700
Parameters Beta 0.0000 0.0000 0.0000
Gamma 0.0000 0.0000
Sum of Squared Residuals 2943240. 3114186. 2932171. 1040425. 1021982.
Root Mean Squared Error 81.51002 83.84370 81.35660 48.46225 48.03080
Mean 1042.763 1117.194 1113.848 1093.138 1094.759
End of Period Levels
Trend 1.345373 1.359729 0.878545 0.878545

Table 7

Daily Data Traffic Estimation Output of AIC and AMSE Based ETS Chosen
Models.

AIC Based AMSE Based

Chosen Models

E E{A} E{A}
T T{N} T{MD}
S S{M} S{M}
Parameters
Alpha 0.669730 0.457427
Beta 0.000000
Gamma 0.000000 0.000000
Phi 0.729664
(continued)

14
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

AIC Based AMSE Based

T T{N} T{MD}
S Estimation information
S{M} S{M}
Parameters
AIC 6147.985 6177.137
Alpha
SSR 0.669730 0.457427
1022170. 1077010.
BetaRMSE 0.000000
48.03521 49.30693
Gamma
AMSE 0.000000 0.000000
3448.010 3326.265
Phi 0.729664
Estimation information
AIC 6147.985 6177.137
SSR
Table 8 1022170. 1077010.
RMSE 48.03521 49.30693
AMSE 3448.010 3326.265
Comparison of HWMS and ETS Methods in Case of Daily Data Traffic.
Table 8
Simple
Comparison of HWMS and ETS Methods in Case of Daily Data Traffic. ETS
Exponential Smoothing Methods HWMS {A,N,M}
Simple ETS
Exponential SmoothingAlpha
Methods
HWMS {A,N,M} 0.6700 0.669730
Beta 0.669730 0.0000
ParametersAlpha 0.6700
Beta Gamma 0.0000 0.0000 0.000000
Parameters
GammaPhi 0.0000 0.000000
Akaike Information
Phi Criterion 6147.985
SumInformation
Akaike of Squared Residuals
Criterion 6147.985 1021982. 1022170.
SumRoot Mean Residuals
of Squared Squared Error 1021982. 1022170. 48.03080 48.03521
Average Mean Squared Error48.03080
Root Mean Squared Error 48.03521 3448.010
Average Mean Squared Error 3448.010

Figure 7. In-sample forecast by HWMS and ETS in case of daily data

Figure 7. In-sample forecast by HWMS and ETS in case of daily data traffic.
traffic.

15 12
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

Figure 8. In-sample
Figure forecast by
8. In-sample HWMS by
forecast and HWMS
ETS for 3and
weeks
ETSin case
for of3 daily
weeks dataintraffic.
case of
daily data traffic.

Figure 8. In-sample forecast by HWMS and ETS for 3 weeks in case of daily data traffic.

Figure 9. Out-of-sample forecast by HWMS and ETS for 3 weeks in case of daily data traffic.

CONCLUSIONS

In network planning and management, short-term prediction and real-time decisions are best served
Figure 9. Out-of-sample
Figure forecast byforecast
9. Out-of-sample HWMS and ETS for 3and
by HWMS weeks
ETSin case
for of daily data
3 weeks intraffic.
case
by simple and low-cost prediction methods with a high accuracy. Therefore, exponential smoothing
of daily data traffic.
methods, such as HWMS, were applied to forecast GSM/GPRS network Erlang traffic. However, a
CONCLUSIONS
more comprehensive study was required to evaluate the usage of various exponential smoothing
In network planning and management, short-term prediction and real-time decisions are best served
methods in more types of wireless traffic, such as
16adata, which is becoming more important along with
by simple and low-cost prediction methods with high accuracy. Therefore, exponential smoothing
the development of communication technology. This research applied different exponential smoothing
methods, such as HWMS, were applied to forecast GSM/GPRS network Erlang traffic. However, a
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

CONCLUSION

In network planning and management, short-term prediction and real-time

decisions are best served by simple and low-cost prediction methods with a
high accuracy. Therefore, exponential smoothing methods, such as HWMS,
were applied to forecast GSM/GPRS network Erlang traffic. However, a more
comprehensive study was required to evaluate the usage of various exponential
smoothing methods in more types of wireless traffic, such as data, which is
becoming more important along with the development of communication
technology. This research applied different exponential smoothing methods
that were categorized as a simple exponential smoothing method and ETS
framework, to forecast cellular network traffic that included voice and data.
The experiments on hourly and daily traffic collected from commercial cellular
networks showed that HWMS was the best fit for cellular network traffic
among other simple exponential smoothing methods that included single,
double, Holt-Winters No Seasonal, and Holt-Winters Additive Seasonal. It was
unsuitable for the ETS framework to forecast hourly voice traffic. However,
in the case of daily voice and data traffic, the ETS found that the {A,N,M}
models provided nearly the same results as the HWMS. The HWMS forecast
series was close to the original series in both hourly and daily voice traffic
cases and showed a good ability to forecast daily data traffic as well. Due to
the low complexity and low cost, HWMS can be applied effectively to cellular
network traffic prediction. Moreover, the evaluation of HWMS can be further
implemented based on the specific requirements of cellular networks.

Acknowledgment

This study was supported by National Natural Science Foundation of China

(NSFC No.61271245).

REFERENCES

5G Vision. (2015). Retrieved from https://5g-ppp.eu/roadmaps

Fletcher, R. (1987).  Practical methods of optimization  (2nd ed.). New
York: John Wiley & Sons.
Hyndman, R. J., Koehler, A. B., Snyder, R. D., & Grose, S. (2002). A state
space framework for automatic forecasting using Exponential smoothing
methods. International Journal of Forecasting, 18(3), 439–454. doi:
10.1016/S0169-2070(01)00110-8
Margaret, V., & Jose, J. (2015). Exponential smoothing models for prediction
of solar irradiance. International Journal of Advanced Research in

17
 Journal of ICT, 18, No. 1 (January) 2019, pp: 1–18

Electrical, Electronics and Instrumentation Engineering, 4 (2), 1133-

1139. doi: 10.15662/ijareeie.2015.0402083
Matta, P., Pant, B., & Arora, M. (2017). All you want to know about Internet of
Things (IoT). International Conference on Computing, Communication
and Automation (ICCCA2017). Greater Noida, India: IEEE. doi:
10.11.09/CCAA.2017.8229999
NGMN 5G White Paper. (2015). Retrieved from https://www.ngmn.org/5g-
white-paper/5g-white-paper.html
Qiang, L., Li, J., & Altman, E. (2017). A novel distributed network selection
scheme for heterogeneous wireless network environments. IEEE
Transactions on Control of Network Systems, 4(3), 575-586. doi:
10.1109/TCNS.2016.2541338
Rahman, M. H., Salma, U., Hossain, M. M., & Khan, M. T. F. (2016). Revenue
forecasting using holt–winters exponential smoothing. Research &
Reviews: Journal of Statistics, 5(3), 19-25.
Siregar, B., Butar-Butar, I. A., Rahmat,. R.F., Andayani, U., & Fahmi, F.
(2017). Comparison of Exponential Smoothing methods in forecasting
palm oil real production. Journal of Physics: Conference Series,801(1),
1-9. doi:10.1088/1742-6596/801/1/012004
Tikunov, D., & Nishimura, T. (2007). Traffic prediction for mobile network
using holt-winter’s exponential smoothing. 15th International
Conference on Software, Telecommunications and Computer Networks,
SoftCOM 2007. Split-Dubrovnik, Croatia: IEEE. doi: 10.1109/
SOFTCOM.2007.4446113
Tran, Q. T., Ma, Z., Li, H., Hao, L., & Trinh, Q. K. (2015). A multiplicative
seasonal ARIMA/GARCH model in EVN traffic prediction.
International Journal of Communications, Network and System
Sciences,8(4), 43-49. doi: 10.4236/ijcns.2015.84005
Tyagi, V., & Kumar, A. (2017). Internet of things and social networks: A
survey. International Conference on Computing, Communication
and Automation. Greater Noida, India: IEEE. doi: 10.1109/
CCAA.2017.8230013
Usaratniwart, E., Sirisukprasert, S., Hatti, N., & Hagiwara, M. (2017). A case
study in micro grid using adaptive enhanced linear exponential
smoothing technique. 8th International Conference of Information
and Communication Technology for Embedded Systems. Chonburi,
Thailand: IEEE. doi: 10.1109/ICTEmSys.2017.7958776
Valakevicius, E., & Brazenas, M. (2015). Application of the seasonal Holt-
Winters model to study exchange rate volatility. Inzenerine Ekonomika-
Engineering Economics,26(4), 384-390.doi: 10.5755/j01.ee.26.4.5210
Yusuf-Asaju, A.W., Dahalin, Z.M., & Ta'a, A (2018) Framework for modelling
mobile networks quality of experience through big data analytics
approach. Journal of ICT, 17(1) 79-113.

18