TELKOM
NIKA
, Vol.11, No
.11, Novemb
er 201
3, pp. 6413
~6
419
e-ISSN: 2087
-278X
6413
Re
cei
v
ed Ma
rch 2
9
, 2013;
Re
vised June
17, 2013; Accepte
d
Jul
y
1
,
2013
A Hybrid Neural Network Prediction Model of Air Ticket
Sales
Han
-
Ch
en H
u
ang
Dep
a
rtment of Leis
u
re Man
a
g
e
ment, Yu Da
Univers
i
t
y
, T
a
iw
a
n
e-mail: hc
hua
n
g
@
y
d
u
.edu.t
w
A
b
st
ra
ct
Air ticket sale
s reven
ue is
a
n
i
m
porta
nt so
urce
of rev
enu
e for travel a
g
enci
e
s, and
if future a
i
r
ticket sal
e
s r
e
venu
e c
an
be
accurate
ly for
e
cast, travel
ag
enci
e
s w
ill
b
e
abl
e to
adv
anc
e pr
ocure
m
ent
to
achi
eve a
sufficient
a
m
ou
nt
of cost-effective tickets. T
her
efore, this stu
d
y ap
pli
ed th
e
Artificial
Neur
a
l
Netw
ork (ANN
) and G
enetic
Algorit
h
m
s (GA) to estab
lish
a pre
d
icti
on
mo
de
l of trave
l
ag
ency
air ti
cke
t
sales r
e
ve
nue.
By verifyi
ng th
e e
m
p
i
rica
l d
a
t
a
, this st
u
d
y pr
oved
that the
e
s
tablis
hed
pre
d
i
ction
mod
e
l
ha
s
accurate
pre
d
i
c
tion
pow
er, a
nd MAPE
(
m
e
an
abso
l
ute
p
e
r
centag
e
error)
is o
n
ly
9.1
1
%
.
T
he esta
blis
h
e
d
mo
de
l ca
n pr
ovid
e b
u
si
nes
s op
erators w
i
t
h reli
ab
le
an
d effici
ent pr
e
d
ictio
n
d
a
ta
a
s
a ref
e
renc
e
for
oper
ation
a
l d
e
c
isions.
Ke
y
w
ords
:
arti
ficial n
eura
l
net
w
o
rk, genetic
a
l
gorit
hms, a
i
r ticket, sales rev
enu
e
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Travel i
s
not
a ne
ce
ssity
in daily livelihood,
thu
s
,
the touri
s
m
indu
stry is
dire
ctly
impacte
d d
u
ri
ng e
c
o
nomi
c
downturn.
Travel ag
en
cies nee
d to
have
goo
d fina
nci
a
l man
age
me
nt
and ri
sk ma
n
ageme
n
t cap
abilities in order to su
rv
ive in the com
petitive market [1]. Financial
manag
eme
n
t is critical to travel agen
cie
s
.
The strengthe
ning o
f
financial m
anag
ement can
increa
se in
come an
d re
duce co
sts.
By analyzing and
add
ressin
g fina
ncial p
r
obl
e
m
s,
comp
etitiveness ca
n be
enha
nced.
Among the bu
sine
ss scop
e
of travel agenci
e
s, air ti
cket
sale
s are an
important source
of rev
enue. Travel
agenci
e
s p
u
rcha
se a la
rge num
be
r of
discou
nted
ai
r ticket
s fro
m
airlin
ers
and
sell
to
con
s
u
m
ers [2]. If a
travel ag
en
cy ca
n a
c
curate
ly
predi
ct market demand, they can pu
rcha
se
suffi
ci
ent numbe
rs of discou
n
ted air ticket
s in
advan
ce to g
a
in high
er p
r
ofits, and re
d
u
ce the
a
c
cu
mulated
cost
s ca
used by over pu
rcha
ses o
r
orde
r l
o
sse
s
cau
s
e
d
by
st
ock d
epletion
[3]. Ther
efore,
this study applie
d
the Artificial Neu
r
al
Network (A
NN) a
nd G
enet
ic Algo
rithms
(GA) to
es
ta
b
lish a
pre
d
icti
on mod
e
l for
air ticket sale
s
r
e
ve
n
u
e
.
T
he r
e
s
e
a
r
c
h
find
in
gs
c
a
n pro
v
id
e
th
e
ind
u
stry with a pra
c
tical
refe
ren
c
e
of
hi
gh
er
reliability and
efficien
cy.
2. Literature
Rev
i
e
w
2.1. Trav
el Agenc
y
Travel
age
nci
e
s
are the
int
e
rme
d
iary
bet
wee
n
tou
r
i
s
m
produ
ct
sup
p
liers a
nd
cu
st
omers,
and a
r
e
re
sp
onsi
b
le for pl
annin
g
an
d a
rra
ngem
ents
of travel for
custome
r
s to
gain
saving
s
[3].
Acco
rdi
ng to
Taiwa
n
’s Stat
ute for the
Developme
n
t o
f
Touri
s
m [4]
Article 2:
“Travel enterpri
s
e:
also
referre
d
to as travel a
gen
cy, a profi
t
-tak
ing
enterprise lice
n
sed
by
the centra
l administ
r
ative
authority to
p
r
ovide to
uri
s
t
s
with a
r
rang
ed travel
sch
edule,
boa
rd
and l
odgi
ng, t
our guid
e
, an
d to
purcha
s
e
tra
n
sp
ortation
tickets
and
ap
ply for tr
avel
document
s a
nd visas on
touri
s
ts’
beh
al
f, a
s
well a
s
to p
r
ovide rel
a
ted
se
rvice
s
for
remu
ner
ation
”
. Taiwan’
s Regulatio
ns
G
o
vernin
g Tra
v
el
Agenci
e
s [5]
Article 3: Tra
v
el agen
cie
s
are divi
de
d i
n
to con
s
oli
d
a
t
ed travel ag
enci
e
s, Cl
ass-A
travel agen
ci
es, and
Cla
s
s-B travel age
ncie
s:
A. The busin
ess scop
e of con
s
oli
dated
travel
agen
ci
es shall con
s
i
s
t of the following:
1)
Being co
mmi
ssi
one
d to se
ll passeng
er t
i
ckets
for do
mestic o
r
foreign land, se
a, and air
transpo
rtation
operators,
or
to buy d
o
m
estic
or
oversea
s
pa
ssenge
r tickets and to
handl
e ship
m
ent of luggag
e for travelers.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 641
3 – 6419
6414
2)
Being commi
ssi
one
d to ha
ndle inb
oun
d
and outb
oun
d pro
c
e
dures and visa
app
lication
s
for travelers.
3)
Soliciting of busi
n
e
ss fro
m
or re
ceivin
g dome
s
tic a
nd oversea
s
tourist
s
, and
arrangi
ng
for their tou
r
s, food, acco
m
m
odation, an
d transpo
rtation.
4)
Arran
g
ing do
mestic and
o
v
erse
as
to
urs,
food and a
c
commo
datio
n, and tran
sp
ortation
for travelers as well as related se
rvices,
thro
ug
h cha
r
tere
d or self
-organ
ized tou
r
packa
ge
s.
5)
Commi
ssioni
ng of Cla
s
s-A
travel agen
ci
es to soli
cit b
u
sin
e
ss.
6)
Commi
ssioni
ng of Cla
s
s-B
travel agen
ci
es to soli
cit d
o
mesti
c
tour
grou
p bu
sine
ss.
7)
Han
d
ling
of communi
catio
n
, prom
otion,
and
pri
c
e q
u
o
tation on
be
half of forei
g
n travel
agen
cie
s
.
8)
Planning
of dome
s
tic an
d
ov
erse
as tours, and
a
rra
ngin
g
for tour
g
u
ide
s
or
tou
r
manag
ers.
9)
Providing of a
d
vice on d
o
m
e
stic a
nd ove
r
se
as tou
r
s.
10)
Operating other dome
s
tic
and overse
as
tou
r
-rel
ate
d
bu
sine
sse
s
as
app
rove
d by the
central admi
n
istrative auth
o
rity.
B. The busin
ess scop
e of Cla
s
s-A trave
l
agen
cie
s
sh
all con
s
i
s
t of the followi
ng:
1)
Being co
mmi
ssi
one
d to se
ll passeng
er t
i
ckets
for do
mestic o
r
foreign land, se
a, and air
transpo
rtation
operators,
or
to buy d
o
m
estic
or
oversea
s
pa
ssenge
r tickets and to
handl
e ship
m
ent of luggag
e for the traveler.
2)
Being commi
ssi
one
d to ha
ndle inb
oun
d
and outb
oun
d pro
c
e
dures and visa
app
lication
s
for travelers.
3)
Soliciting of
busi
n
e
ss fro
m
or
receivin
g
dome
s
tic
a
nd overse
as
tourist
s
; arra
nging fo
r
their tour, foo
d
, accommo
d
a
tion, and tra
n
sp
ortation.
4)
Arran
g
ing
of
oversea
s
tou
r
s, food
an
d a
c
commo
datio
n, and
tra
n
sp
ortation
for t
r
avelers
as well as p
r
o
v
ision of relat
ed se
rvice
s
, throu
gh self-o
rgani
ze
d tour
packa
ge
s.
5)
Soliciting of b
u
sin
e
sse
s
on
behalf of co
n
s
olid
ated trav
el agen
cie
s
.
6)
Planning
of dome
s
tic an
d
ov
erse
as tours, and
a
rra
ngin
g
for tour
g
u
ide
s
or
tou
r
manag
ers.
7)
Providing of a
d
vice on d
o
m
e
stic a
nd ove
r
se
as tou
r
s.
8)
Operating other dome
s
tic
and overse
as
tou
r
-rel
ate
d
bu
sine
sse
s
as
app
rove
d by the
central admi
n
istrative auth
o
rity.
C. The bu
sin
e
ss scop
e of Cla
s
s-B trave
l
agen
cie
s
sh
all con
s
i
s
t of the followi
ng:
1)
Being comm
issi
one
d to sell p
a
ssen
g
e
r tickets fo
r dom
esti
c l
and, sea, a
nd air
transpo
rtation
operators,
or to buy dome
s
tic p
a
sse
n
g
e
r ticket
s and
handl
e ship
ment of
lugga
ge on b
ehalf of travelers.
2)
Soliciting of
b
u
sin
e
ss
from or re
ceiving
dome
s
tic to
urists; a
r
rangi
n
g
for t
ours, fo
od a
nd
accomm
odati
on, and tran
sportation; a
n
d
providing of related services.
3)
Soliciting of
b
u
sin
e
sse
s
rel
a
ted to do
me
st
ic g
r
o
up tou
r
s
on b
ehalf
of con
s
oli
dat
ed travel
agen
cie
s
.
4)
Planning of d
o
mesti
c
tours.
5)
Providing of a
d
vice on d
o
m
e
stic tou
r
s.
6)
Operation of
other dome
s
tic tour-rela
t
ed busi
nesses a
s
app
ro
ved by the
central
admini
s
trativ
e authority.
2.2. Artificial
Neural Net
w
ork
ANN is a ma
thematical m
odel for sim
u
lati
ng the structure and fu
nction of a biologi
cal
neural net
wo
rk. An A
NN
condu
cts
co
m
putation by la
rge
numb
e
rs
of artificial
ne
uron
s [6, 7]. I
n
most ca
se
s, ANN can cha
nge an inte
rn
al stru
cture a
c
cordi
ng to e
x
ternal inform
ation as a typ
e
of
adaptive
syst
em [6-8]. ANN is
a no
n-lin
ear
statis
ti
cal
data mod
e
lin
g tool commo
nly use
d
for t
he
model con
s
truction of com
p
lex input-o
utput relation
sh
ips or the exp
l
oration of a d
a
ta model [9].
The ANN
co
nstru
c
tion
pri
n
cipl
e is
origi
nat
ed by the
ope
ration
of biologi
cal
(h
uman
o
r
other a
n
imal
s) n
eural ne
tworks. A
N
N, like
hum
a
n
s, ha
s si
m
p
le de
cisi
on
and jud
g
m
ent
capabilities, whi
c
h
are more
adv
antageous
as compared with form
al logi
cal
reasoning [7-10].
A
comm
on mult
ilayer feedfo
r
ward network con
s
ist
s
of three pa
rts
(Fig
ure 1
)
[11, 12
]:
1)
Input layer, a larg
e num
ber of ne
uro
n
s fo
r the
rece
ption of
a larg
e amo
unt of input
information.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Hyb
r
id Neu
r
al Net
w
o
r
k P
r
edi
ction Mo
d
e
l of Air Ticke
t
Sales (Han-Che
n
Hu
ang
)
6415
2)
Output layer, the inform
ation is tra
n
sferre
d, an
alyzed, an
d
weigh
ed in
the neuron
c
o
nnec
tions
to form the output results
.
3)
Hidd
en layer,
the layers fo
rmed by the
num
e
r
ou
s ne
uron
s a
nd lin
ks b
e
twe
en t
he input laye
r
and outp
u
t layer. It can be of multipl
e
layers
, b
u
t is usu
a
lly only one layer. There is n
o
recogni
ze
d st
anda
rd for th
e numb
e
r of
neuron
s in
th
e hidde
n laye
r. Ho
wever, n
e
tworks with
more
neu
ron
s
a
r
e m
o
re
si
gnifica
ntly no
n-line
a
r,
a
nd thus,
the neu
ral
n
e
two
r
k ro
bustn
ess will
be more sig
n
i
f
icant.
The
Ba
ck
Propag
ation Ne
ural Net
w
ork (BPNN)
i
s
th
e cu
rrently th
e mo
st re
pre
s
entative
and
mo
st co
mmonly u
s
e
d
ANN [11
-
13]
. BPNN u
s
e t
he
steep
est
desce
nt meth
od to
adju
s
t t
h
e
netwo
rk p
a
ra
meters to determin
e
more accu
rate sol
u
tions by iterat
ive computin
g.
Figure 1. BPNN Archite
c
t
u
re
2.3.
Genetic Algorithms (GA)
The GA i
s
a
rand
om
sea
r
ch meth
od th
at simu
late
s t
he survival of
the fittest biologi
cal
evolutiona
ry law, a
s
p
r
op
ose
d
by Joh
n
He
nr
y
Holl
and [14]. Wi
th the se
arch algo
rithm
of
“su
r
vival an
d
detectio
n
”,
it has b
een
widely u
s
e
d
in combin
atorial o
p
timi
zation, ma
ch
ine
learni
ng, sig
nal pro
c
e
s
si
ng, self-a
dap
tive c
ontrol, and othe
r fields. GA is one of the key
techn
o
logie
s
relating to i
n
telligent comp
uting of the
pre
s
ent a
nd
has
been
ap
plied by ma
n
y
enterp
r
i
s
e
s
for timetable
arrang
ement
s, data anal
y
s
is, future tre
nd pre
d
ictio
n
, budgeting,
and
solving ma
ny other combi
n
atoria
l optimi
z
ation p
r
obl
e
m
s [15-17].
In GA, the
so
lution to
a p
r
oblem i
s
call
ed a
n
in
dividual, an
d u
s
e
s
a va
riable
seque
nce
(kn
o
wn a
s
chrom
o
some
). The
chrom
o
som
e
i
s
ge
nerally exp
r
e
s
sed a
s
a
simple
string
or
nume
r
ic
stri
n
g
. Firstly, the algo
rithm ra
ndomly
g
ene
rates a
ce
rta
i
n num
ber of
individual
s.
In
each g
ene
rati
on, ea
ch
indi
vidual i
s
eval
uated
and
a fi
tness valu
e i
s
obtain
ed
by the
com
putation
of the fitness
function. Indi
vidual
s of the
populatio
n will be so
rted a
c
cordi
ng to fitness value
s
i
n
desce
nding
o
r
de
r. The n
e
xt step is to g
e
nerate
th
e ne
xt generatio
n
of individual
s, and thu
s
, the
formation
of t
he p
opul
ation
.
This p
r
o
c
e
s
s i
s
im
pleme
n
ted via
sele
ction,
crosso
ver, an
d m
u
tation
[16, 18]. When the fitness value is hi
gher, the op
portunity of b
e
ing sele
cted
will be high
er. A
relatively opti
m
ized
po
pula
t
ion is form
e
d
thro
ugh
s
u
ch
a
sele
ct
io
n me
cha
n
is
m
.
A
f
t
e
rwa
r
ds,
t
he
sele
cted
indiv
i
dual
s u
nde
rg
o the
process of cro
s
so
ver.
In ge
ne
ral,
GA ha
s
a ran
ge of
cro
s
sover
probabilities, being
0.6-1.
Furthermore,
it is the
acti
on of
mutation, through
which
new “sub-
individual
s”
a
r
e g
ene
rated.
In gene
ral,
GA ha
s a fi
xed mutatio
n
con
s
tant of
0
.
1 or b
e
lo
w. After
sele
ction, cro
s
sover, an
d mutation, the
best indi
vidu
al has m
o
re
oppo
rtunitie
s
of being sele
cted
to gene
rate t
he next ge
neration, whil
e in
dividual
s of
lo
wer fitne
s
s va
lue are g
r
adu
ally eliminate
d
.
Such a
process is
rep
e
a
t
ed until the
endin
g
c
o
n
d
itions
are
satisfied. The
gene
ral e
n
d
i
ng
con
d
ition
s
are as follo
ws [
15-1
9
]:
1)
Limited evolu
t
ionary times;
2)
Comp
uting reso
urce con
s
traints (su
c
h
as
the comp
utation time and mem
o
ry
occupied by
comp
utation
)
;
3)
An individual
has met the
condition
s of the optimal va
lue;
4)
Contin
ued ev
olution do
es
not prod
uce
individual
s wit
h
better fitness value
s
;
.
.
Out
put
W
ij
Input
Input
la
y
e
r
Hidde
n La
y
e
r
Out
put
la
y
e
r
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 641
3 – 6419
6416
5) Huma
n
interv
ention;
6)
Combi
nation
s
of two or more of
the situations d
e
scri
be
d in the abov
e.
3. Rese
arch
Metho
d
3.1. Input Variables
Based
on lite
r
ature review [20-2
6
], this
study
u
s
ed
in
ternation
a
l oil
pri
c
e, Tai
w
a
n
sto
c
k
market
weig
h
t
ed ind
e
x, NTD/US
D ex
chang
e rate,
peopl
e traveli
ng a
b
ro
ad f
r
om Tai
w
a
n
e
a
ch
month, Taiwan’s m
onthly unemploym
ent rate, Ta
i
w
an’
s monthl
y monitor ind
i
cator,
T
a
i
w
a
n
’
s
monthly co
mposite
lea
d
ing inde
x,
T
a
i
w
a
n
’
s
m
o
n
t
h
l
y
c
o
mposite coincid
ent inde
x,
a
n
d
K
travel ag
en
cy’s mo
nthly ai
r ticket sale
s revenue
(T
-1
month~T-1
2
month) (T
abl
e 1),
as the i
nput
variable
s
for
the predi
ction
of K travel agen
cy’s
air ticket sale
s re
venue in T month. The time
perio
d of dat
a sele
ction i
s
from Ja
nua
ry 2004 to
Au
gust 2
012. M
o
reove
r
, 68%
data is rand
omly
sele
cted a
s
the trainin
g
da
ta, 16% as the cro
s
s
valida
t
ion data, and
16% as the testing d
a
ta. GA
improve
s
the
perfo
rma
n
ce
of ANNs by
sele
cti
ng th
e
optimum i
n
p
u
t feature
s
.
This
study
u
s
ed
different o
p
e
r
ators for sele
ction
(Roulett
e
, Best, Rand
om, and
Tou
r
nament
) a
nd
cro
s
sove
r (O
ne
point, Uniform, Arithmetic, and He
uri
s
tic) ope
ration
s.
Table 1. Fo
re
ca
st Model V
a
riabl
es
Variables
Input
International oil price (T-1 m
onth)
Taiwan stock market
w
e
ighted ind
e
x(
T-1 mo
nth)
NTD/US
D excha
nge rate
(T-
1
mo
nth)
People traveling abroad f
r
om Tai
w
a
n
each month
(T-1 m
onth)
Taiwan’s monthly unemplo
y
ment
rate(
T
-1 mo
nth)
Taiwan’s monthly monitor indicat
o
r(T
-1 mont
h)
Tai
w
a
n
’s
mon
t
hl
y co
mp
osite
l
eadi
ng
in
de
x
(
T
-
1
m
o
n
t
h
)
Taiwan’s monthly co
m
p
o
s
i
t
e
c
o
i
n
c
i
d
e
n
t
i
n
d
e
x
(
T
-1 mont
h)
K travel agenc
y
’
s air ticket sales r
e
venue (T
-1 mon
t
h to T-1
2
month
)
Output
Air ticket sale
s revenue(T mo
nth)
3.2. Architec
ture Design
and Model T
r
aining
Reg
a
rdi
ng A
NN, the in
put
activation fu
ncti
on u
s
e
s
Hyperboli
c
T
ange
nt, the output error
function u
s
e
s
Sum-Of-S
quares, an
d
the output
activation functio
n
use
s
Logi
stic.
The
architectu
re
desi
gn uses Test
E
rro
r as
the
ju
dgm
en
t stand
ard
when
se
archin
g for the
opti
m
al
netwo
rk a
r
chi
t
ecture
(Figu
r
e 2).
Figure 2. Best Netwo
r
k Architecture Sea
r
ch
Re
sults
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Hyb
r
id Neu
r
al Net
w
o
r
k P
r
edi
ction Mo
d
e
l of Air Ticke
t
Sales (Han-Che
n
Hu
ang
)
6417
ANN T
r
ainin
g
Algorithm: Q
u
ick Prop
agat
ion Algorithm,
Trainin
g
Algorithm’
s
Para
meters
inclu
de: Qui
c
k Prop
agatio
n Coeffici
ent
= 1.75, Lea
rn
ing Rate
= 0.1
.
Overtrainin
g
control a
nd the
weig
hts ra
nd
omizatio
n me
thod are u
s
e
d
to increa
se
model a
c
cura
cy (Figu
r
e 3
)
.
Figure 3. Net
w
ork T
r
ainin
g
Options
4. Empirical
Resul
t
s
Mean absolut
e
percentage error (MAPE) and
co
rrelat
ion (r) are adopted as indi
cators
for evaluating
the model.
1)
MAPE: The smaller the val
ue, the
small
e
r the
error between the foreca
st val
ue
and the target
value.
%
100
'
1
1
n
i
i
i
y
y
y
n
MAPE
(1)
Whe
r
e
n i
s
the n
u
mb
er o
f
the fo
re
cast
ing p
e
ri
ods,
y
i
is the actual value for the i
perio
d, and
i
y
'
is the fore
ca
st
value for the i period.
2)
Correl
ation (r): As r app
roa
c
he
s 1,
the ANN forecastin
g results imp
r
ove.
The actu
al values an
d mod
e
l output values dist
ributio
ns are as sho
w
n in Figu
re 4. Most
of the output values a
r
e
distrib
u
ted al
ong both
sid
e
s of the dia
gonal lin
e (O
utput/Targ
e
t=1),
indicating tha
t
the model has go
od predi
ctability.
Figure 4. Sca
tter Plot of Actual Value an
d
Model Outp
ut Value
Figure 5. Actual Value an
d
Model Outpu
t
Value Graph
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 641
3 – 6419
6418
Table 2. Mod
e
l Traini
ng Result
s
Target
O
u
tput
Absolute
Error
MAPE
Mean
2771725.58
2800445.86
297705.03
0.0894
SD
575737.53
541518.39
232589.28
0.1048
Correlation:0.
87
The a
c
tual v
a
lue an
d mo
del output v
a
lue tre
n
d
s
are a
s
sho
w
n in Figu
re
5. The
predi
ction
m
odel e
s
tabli
s
hed fo
r ai
r ti
cket sale
s re
venue in
this study h
a
s g
ood
cap
abilit
y to
reflect th
e ch
ange
s in
air ti
cket sal
e
s. T
he re
sult
s
of t
he mod
e
l trai
ning a
r
e a
s
shown in T
abl
e 2.
Mean absol
u
te percentage
error (M
APE) is 8.94%, Correl
a
tion Coef
fici
ent is 0.87,
indicating that
the traini
ng
model
ca
n le
arn th
e trai
ni
ng data
to
a l
e
vel of erro
r
belo
w
10%.
The mo
del te
sting
results
are
a
s
sho
w
n in
T
able 3, MAP
E
is 9.11%
,
Correl
ation
Coefficient i
s
0
.
83, rep
r
e
s
en
ting
that the p
r
edi
ction m
odel
h
a
s th
e a
b
ility for a
c
cura
te
predi
ction
of
air ticket
sal
e
s revenu
e. T
he
use
r
interfa
c
e
of the establi
s
he
d model i
s
as
sho
w
n in
Figure 6.
Table 3. Mod
e
l Fore
ca
st Result
s
Target
O
u
tput
Absolute
Error
MAPE
Mean
2988092.89
2835198.96
331872.74
0.0911
SD 625322.11
422310.58
277259.95
0.1169
Correlation:0.
83
Figure 6. The
Use
r
Interfa
c
e of the Establish
ed Mod
e
l
5. Conclusio
n
This stu
d
y used BPNN a
n
d
GA to establish a
predi
ctio
n model for travel agen
cie
s
on air
ticket sale
s revenue. Emp
i
rical
re
sults
showed t
hat, the pro
p
o
s
ed
model ha
s go
od pre
d
icta
bil
i
ty
and the p
r
edi
ction mo
del
has the
abilit
y of accu
ra
te
ly predi
cting
air ticket sal
e
s reve
nue. T
h
e
MAPE is 9.11%, and Correlation C
oefficient is
0.83. If a travel
agency can predict the changes
in air ticket
sale
s
reven
u
e
in
advan
ce, it ca
n p
u
rcha
se
di
scou
nted ai
r tickets in
suffici
ent
numbe
rs. Th
e propo
se
d m
odel
can
p
r
ovide bu
sin
e
ss
with a
more reliable
and
ef
ficient referen
c
e
basi
s
in practi
ce.
Referen
ces
[1]
Z
hu Z
,
Z
hao J. Comparativ
e
Stud
y
on the
Sources of e-
Busin
e
ss Com
petitive Adv
ant
ages b
e
t
w
e
e
n
T
r
avel Agencie
s and Onli
ne T
r
avel Servic
e F
i
rms.
Journal o
f
China Un
iver
sity of Geoscie
n
ces Soci
a
l
Scienc
es Editi
o
n
. 201
1; 11(2)
: 114-11
9.
[2]
Che
n
SY, Liu
HH. A Stud
y
on Busi
ness
Mode
ls of
On-Lin
e T
r
avel Agenc
y-T
he Cas
e
of Eztravel.
Journ
a
l of Le
is
ure an
d T
ouris
m Ind
u
stry Res
earch
. 20
10; 5(
1): 90-99.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Hyb
r
id Neu
r
al Net
w
o
r
k P
r
edi
ction Mo
d
e
l of Air Ticke
t
Sales (Han-Che
n
Hu
ang
)
6419
[3]
Dai MH. U
n
fair
Competiti
on
in
T
r
avel Service
Industr
y
.
J
our
nal
of Eastern
Lia
oni
ng U
n
iv
e
r
sity (Social
Scienc
es)
. 201
1; 13(2): 38-4
2
.
[4]
Yang CK. T
ourism Administra
tion an
d La
w
s
.
T
a
ipei: Yang C
h
ih Ltd. 2
012.
[5]
Niu HY. T
r
avel Industr
y
M
ana
geme
n
t.
T
a
ipei
: Yang Chi
h
Lt
d. 2005.
[6]
Guo H, Neur
al Net
w
o
r
k Aide
d Kalma
n
F
ilter
ing for I
n
tegrate
d
GPS/INS Navigat
ion S
y
stem.
TEL
K
OMNIKA
. 2013; 1
1
(3): 1
221-
122
6.
[7]
Hou Y, Mai Y.
Chaotic Pre
d
i
c
tion for T
r
affi
c F
l
o
w
of Improved BP N
eur
al Net
w
o
r
k.
TEL
K
OMNIKA
.
201
3; 11(3): 16
83-1
690.
[8]
Uncin
i
A, Vecci L, Campol
uc
ci P. Piazza F
.
Comple
x-V
a
l
u
ed Ne
ural N
e
tw
o
r
ks
w
i
th Ad
aptive Sp
lin
e
Activation Functions.
IEEE Tr
ansactions on Signal Process
i
ng
. 19
99; 47(
2
)
: 505-51
4.
[9]
Elha
di GF
, F
a
rouk RM, Issa
AT
. Protein
Sequ
enc
e for
Clusteri
n
g
DN
A bas
ed
on
A
r
tificial
Neur
al
Net
w
orks.
Inter
natio
nal J
ourn
a
l of Co
mp
uter
Science Issue
s
. 2012; 9(1): 1
61-1
67.
[10]
You C, H
ong
D
.
Nonli
n
e
a
r Bli
nd Eq
ual
izati
o
n
Schem
es usi
ng C
o
mpl
e
x-V
a
lu
ed Mu
ltila
ye
r F
eedfor
w
a
rd
Neur
al Net
w
o
r
ks.
IEEE
Transactions on Neural Networks
. 1998; 9: 14
42-1
455.
[11]
Hua
ng H
C
, Ch
ang AY, H
o
C
C
. Usin
g Artific
i
al N
eur
al N
e
tw
o
r
ks to Estab
lish
a Custom
e
r-Cance
l
l
a
tio
n
Predicti
on Mod
e
l.
Pr
z
e
gl
ad El
ektrotechn
ic
z
n
y.
2013; 89(
1b)
: 178-18
0.
[12]
Cha
ng
HT
, C
hen
PC, Hu
an
g HC,
Lin
DH
. A Stud
y
o
n
the Ap
plic
ation
of Ne
ural
Ne
t
w
ork
to t
h
e
Predicti
on of
W
e
ight Co
ntrol
.
Internation
a
l
Journ
a
l of Eng
i
neer
ing R
e
se
a
r
ch and D
e
ve
l
o
p
m
e
n
t
. 2013;
5(11): 78-
85.
[13]
Hua
ng HC, H
o
CC. Back-Pro
pag
atio
n Neur
al Net
w
o
r
k Co
mbin
ed
w
i
th
a Particle S
w
a
r
m
Optimization
Algorit
hm for
T
r
avel Pack
age
Dem
a
n
d
F
o
recasti
ng.
Internati
o
n
a
l
Jour
nal
of
Digita
l
C
onte
n
t
T
e
chno
logy
an
d its Applic
atio
ns
. 2012; 6(
17)
: 194-20
3.
[14]
Holl
an
d JH. Adaptatio
n in N
a
tural a
nd Artifici
al
S
y
st
ems. Miami: T
he Unive
r
sit
y
of Michi
g
a
n
. 1975.
[15]
Che
n
Y, Ma
Y
J
, Yun W
X
. A
pplic
atio
n
of Improv
ed
Ge
neti
c
Al
go
ri
th
m
in
PI
D Controller
Par
a
meters
Optimization.
TEL
K
OMNIKA
. 201
3; 11(3): 15
24-1
530.
[16]
Boume
d
ie
ne
L
,
Gao Z
,
Liu S
,
Leg
hmizi S,
Yang
R. Gen
e
tic Alg
o
rithm-B
ased A
ppr
oac
h to Sp
ectru
m
Alloc
a
tion a
n
d
Po
w
e
r Co
ntrol
w
i
th Co
nstraints in C
ogn
itive Ra
dio N
e
tw
o
r
ks.
Research Journal
of
Appl
ied Sci
enc
es, Engin
eer
in
g and T
e
ch
no
l
ogy
. 201
3; 5(1)
: 100-10
7.
[17]
Guan
X,
Men
g
G, Z
h
a
o
C.
Man
u
facturi
n
g R
e
sourc
e
s
Classific
a
tio
n
Based
o
n
F
u
z
z
y
Clust
erin
g
Algorit
hm.
Jour
nal of T
heor
eti
c
al an
d App
lie
d Information T
e
chn
o
lo
gy.
201
2; 46(2): 63
2-6
37.
[18]
Rahm
ani
R, O
t
hman MF
, Yu
sof R, Kh
ali
d
M.
Solvin
g Ec
onom
ic Dis
pat
ch Pro
b
lem
u
s
ing
Particl
e
S
w
a
rm Optimiz
a
tion
b
y
a
n
Ev
oluti
onar
y T
e
chni
que F
o
r Initi
a
lizi
ng P
a
rticle
s.
Journa
l of T
heor
etical
an
d
Appl
ied Infor
m
ation T
e
ch
no
lo
gy
. 2012; 4
6
(2)
:
526-53
6.
[19]
Peng J, Hu
ang
D, W
ang Z
.
Profit-Driven D
y
namic Sp
ectru
m
Allocati
on in
W
i
reless Net
w
orks.
Journa
l
of T
heoretica
l
and Ap
pli
ed Inf
o
rmatio
n
T
e
ch
nol
ogy
. 20
12; 46(2): 61
5-6
1
8
.
[20]
Cha
n
YM. F
o
recastin
g T
ourism: A Sine
W
a
ve T
i
me Seri
es Re
gressi
on
Appro
a
ch.
Jo
urna
l of T
r
avel
Research
. 20
1
1
; 32(1): 58-6
0
.
[21]
W
ang CH. Pre
d
ictin
g
T
ourism Deman
d
usi
ng
F
u
zz
y
T
i
me Series an
d H
y
br
id Gre
y
T
heor
y
.
Tou
r
ism
Mana
ge
me
nt
. 200
4; 25(1): 36
7-37
4.
[22]
Chu F
L
. F
o
rec
a
sting T
ourism
:
A Combine
d
Appro
a
ch.
T
o
u
r
ism Ma
na
ge
ment
. 199
8; 19(6
)
: 515-52
0.
[23]
Chai
bo
onsri
C,
Cha
i
tip P. T
he Mod
e
li
ng Int
e
rnat
i
ona
l T
ourism Dema
nd f
o
r Le
ngth
of s
t
a
y
in In
dia:
Socia
l
Deve
lop
m
ent, Economi
cs Devel
opme
n
t.
Journal of
Know
led
ge Ma
nag
e
m
ent, Econo
mics a
n
d
Information T
e
chno
logy
. 2
012
; 2(2): 81-98.
[24]
Lin CJ, Che
n
HF
, Lee T
S
. Forecasti
ng T
ourism Dema
n
d
usin
g T
i
me Series, Artificial N
eura
l
Net
w
orks
and M
u
ltiv
aria
te Ada
p
tive R
egress
i
on
Spl
i
nes: Evid
enc
e
from T
a
i
w
a
n
.
Internatio
nal
Journ
a
l
of
Business Adm
i
nistration
. 201
1; 2(2): 14-24.
[25] Samsud
in
R, S
aad
P, Sh
abr
i
A. H
y
bri
d
izi
n
g
GMDH a
nd
Le
ast Squ
a
res
S
V
M Sup
port V
e
ctor Mac
h
in
e
for F
o
recasting
T
ourism Dem
and.
Intern
atio
nal Jo
urna
l of Rese
arch an
d Review
s in Ap
plie
d Scie
nces
.
201
0; 3(3): 274
-279.
[26] Petrevska B. Forecasting
International T
ourism Demand:
The Evidence of
Macedonia.
U
T
MS Journa
l
of Econo
mics
.
201
2; 3(1): 45-
55.
Evaluation Warning : The document was created with Spire.PDF for Python.