Khashei, Mehdi; Haji ahimi, Zah a
A icle
Pe o mance e alua ion o se ies and pa allel s a egies
o inancial ime se ies o ecas ing
Financial Inno a ion
P o ided in Coope a ion wi h:
Sp inge Na u e
Sugges ed Ci a ion: Khashei, Mehdi; Haji ahimi, Zah a (2017) : Pe o mance e alua ion o se ies
and pa allel s a egies o inancial ime se ies o ecas ing, Financial Inno a ion, ISSN 2199-4730,
Sp inge , Heidelbe g, Vol. 3, Iss. 24, pp. 1-24,
h ps://doi.o g/10.1186/s40854-017-0074-9
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/176464
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h p://c ea i ecommons.o g/licenses/by/4.0/
RESEARCH Open Access
Pe o mance e alua ion o se ies and
pa allel s a egies o inancial ime
se ies o ecas ing
Mehdi Khashei and Zah a Haji ahimi
*
* Co espondence:
[email p o ec ed]
Depa men o indus ial and sys em
enginee ing, Is ahan Uni e si y o
Technology, Is ahan, I an
Abs ac
Backg ound: Imp o ing inancial ime se ies o ecas ing is one o he mos challenging
and i al issues acing nume ous inancial analys s and decision make s. Gi en i s di ec
impac on ela ed decisions, a ious a emp s ha e been made o achie e mo e accu a e
and eliable o ecas ing esul s, o which he combining o indi idual models emains a
widely applied app oach. In gene al, indi idual models a e combined unde wo main
s a egies: se ies and pa allel. While i has beenp o en ha heses a egiescanimp o e
o e all o ecas ing accu acy, he li e a u e on ime se ies o ecas ing emains ague on
he choice o an app op ia e s a egy o gene a e a mo e accu a e hyb id model.
Me hods: The e o e, his s udy’s key aim is o e alua e he pe o mance o se ies and
pa allel s a egies o de e mine a mo e accu a e one.
Resul s: Acco dingly, he p edic i e capabili ies o i e hyb id models a e cons uc ed on
he basis o se ies and pa allel s a egies compa edwi heacho he andwi h hei base
models o o ecas s ock p ice. To do so, au o eg essi e in eg a ed mo ing a e age
(ARIMA) and mul ilaye pe cep ons (MLPs) a e used o cons uc wo se ies hyb id
models, ARIMA-MLP and MLP-ARIMA, and h ee pa allel hyb id models, simple a e age,
linea eg ession, and gene ic algo i hm models.
Conclusion: The empi ical o ecas ing esul s o wo benchma k da ase s, ha is, he
closing o he Shenzhen In eg a ed Index (SZII) and ha o S anda d and Poo ’s500
(S&P 500), indica e ha al hough all hyb id models pe o m be e han a leas one o
hei indi idual componen s, he se ies combina ion s a egy p oduces mo e accu a e
hyb id models o inancial ime se ies o ecas ing.
Keywo ds: Se ies and pa allel combina ion s a egies, Mul ilaye pe cep ons,
Au o eg essi e in eg a ed mo ing a e age, Financial ime se ies o ecas ing,
S ock ma ke s
Backg ound
Real ime se ies o ecas ing wi h a high deg ee o accu acy is gaining inc easing
impo ance in many domains, pa icula ly he inancial ma ke s, and hus, a ious
a emp s ha e been made o de elop mo e accu a e echniques. The objec i e o
inancial ime se ies o ecas ing is o p o ide inancial analys s and in es o s wi h
eliable guidance on asse managemen . Thus, imp o ing o ecas ing accu acy and
in oducing eliable o ecas ing me hods can acili a e mo e p o i able inancial ma ke
in es men s by lead in es o s and inancie s. To his e ec , choosing a me hod ha
Financia
l
Inno a ion
© The Au ho (s). 2017 Open Access This a icle is dis ibu ed unde he e ms o he C ea i e Commons A ibu ion 4.0 In e na ional
License (h p://c ea i ecommons.o g/licenses/by/4.0/), which pe mi s un es ic ed use, dis ibu ion, and ep oduc ion in any medium,
p o ided you gi e app op ia e c edi o he o iginal au ho (s) and he sou ce, p o ide a link o he C ea i e Commons license, and
indica e i changes we e made.
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24
DOI 10.1186/s40854-017-0074-9
pe o ms well in inancial ime se ies o ecas ing is impe a i e. To p o ide mo e accu -
a e esul s, s udies on ime se ies o ecas ing and modeling widely use a combina ion
o di e en models and me aheu is ic op imiza ion app oaches. Conside able esea ch
has adop ed op imiza ion me hods such as gene ic algo i hm (GA; Aghay Kaboli e al.
2016a, 2016b, Kaboli e al. 2016), pa icle swa m op imiza ion (PSO; Aghay Kaboli e
al. 2016a, 2016b), and gene exp ession p og amming (GEP; Aghay Kaboli e al. 2017a,
2017b; Aghay Kaboli e al. 2016a, 2016b). Kaboli e al. (2016) p oposed he a i icial co-
ope a i e sea ch (ACS) algo i hm o o ecas long- e m elec ici y ene gy consump ion
and nume ically con i med he e ec i eness o he algo i hm using o he me aheu is ic
algo i hms including he GA, PSO, and cuckoo sea ch. Modi i-Delshad e al. (2016)
p esen ed a back acking sea ch algo i hm (BSA) and e i ied he eliabili y o he
me hod in sol ing and modeling he economic dispa ch (ED) p oblem. Aghay Kaboli e
al. (2016a, 2016b) es ima ed elec ici y demand using GEP, a gene ic-based me hod, as
an exp ession-d i en app oach and showed ha GEP ou pe o ms he mul ilaye pe -
cep on neu al (MLP) ne wo k and mul iple linea eg ession models. Recen s udies
on ime se ies o ecas ing la gely ocus on combina ion me hods gi en he
dis inguishing ea u es o hyb id models (e.g., unique modeling capabili y o each
model), d awbacks in using single models, and he esul an imp o emen s in
o ecas ing accu acy. The key concep o combina ion heo y is employing he unique
me i s o indi idual models o ex ac di e en da a pa e ns. Impo an ly, he li e a u e
con i ms ha no indi idual model can uni e sally de e mine da a-gene a ion p ocesses.
In o he wo ds, all cha ac e is ics o unde lying da a canno be ully modeled by one
model and he e o e, combining di e en models o using hyb id ones helps analyze
complex pa e ns in da a mo e accu a ely and comple ely. Fu he , combining a ious
models simpli ies he selec ion o a model ha is app op ia e o p ocess di e en o ms
o ela ionships in he da a and educes he isk o choosing an ine icien one.
Se e al app oaches ha e been p oposed o combine linea and nonlinea models.
These combina ion me hods a e gene ally di ided in o wo p ima y classes: se ies and
pa allel. In a se ies combina ion me hod, a ime se ies is decomposed in o linea and
nonlinea pa s. Acco dingly, in he i s s age, he model is used o p ocess one ime
se ies componen and hen, he ob ained alues a e used as inpu s o he second
model o analyze ano he componen . On he o he hand, in he pa allel combina ion
me hod, he o iginal da a a e simul aneously conside ed o be inpu s o di e en
models and hen, he linea combina ion o he o ecas ed esul s acili a es inal
hyb id o ecas ing.
The li e a u e on se ies linea o nonlinea combina ion models has d ama ically
expanded since he ea ly wo k o Zhang (2003). Fo ins ance, Pai and Lin (2005)
p oposed a se ies hyb id me hodology o exploi he unique s eng h o au o eg essi e
in eg a ed mo ing a e age (ARIMA) and suppo ec o machines (SVMs) o o ecas
s ock p ice and indica ed ha a hyb id model ou pe o ms i s componen s. Chen and
Wang (2007) cons uc ed a se ies combina ion model ha inco po a es seasonal au o-
eg essi e in eg a ed mo ing a e age (SARIMA) and SVMs o seasonal ime se ies
o ecas ing and achie ed mo e accu a e esul s han bo h componen s. Zeng e al.
(2008) p esen ed a se ies combina ion o he ARIMA and MLP models o p edic
sho - e m a ic low. Thei expe imen al esul s o he eal da ase s indica ed ha
he p oposed hyb id model can be an e ec i e in imp o ing o ecas ing accu acy
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 2 o 24
achie ed by ei he componen . Zhou and Hu (2008) conduc ed expe imen s using a hy-
b id modeling and o ecas ing app oach, which was based on he G ey and Box–Jenkins
ARMA models, and showed ha hei p oposed model had highe o ecas ing p ecision
han i s single componen s. Pao (2009) p oposed a hyb id se ies model inco po a ing
a i icial neu al ne wo k (ANN) and di e en ypes o gene alized au o eg essi e
condi ional he e oscedas ici y (GARCH) models o o ecas ene gy consump ion. Table 1
lis s o he ecen s udies on se ies linea and nonlinea hyb id models.
Ba es and G ange (1969) in oduced he concep o a pa allel combina ion, which
was subsequen ly used by many esea che s such as Mak idakis e al. (1982), G ange
and Ramana han (1984), Bunn (1989), and De Menezes and Bunn (1993). Wedding and
Cios (1996) p oposed a pa allel combina ion model using adial basis unc ion
ne wo ks and he Box–Jenkins ARIMA model. Mo e ecen ly, se e al pa allel hyb id
o ecas ing models ha e been p oposed o combine linea and nonlinea models. Fo
ins ance, Wang e al. (2012) p esen ed a pa allel hyb id model using GA and by
employing ARIMA, exponen ial smoo hing (ES), and back p opaga ion neu al ne wo k
(BPNN) models. Thei nume ical esul s showed ha he p oposed model ou pe o ms
all adi ional models, including he ESM, ARIMA, BPNN, equal weigh hyb id (EWH)
model, and andom walk (RWM) model. Fo ecas ing s ock e u ns, Ra he e al. (2015)
p oposed a no el hyb id model ha me ges p edic ions by h ee indi idual models: ES,
ecu en neu al ne wo k (RNN), and ARIMA; he op imum weigh s o each model a e
iden i ied using GA. Yang e al. (2016) p esen ed a combined o ecas ing model using
BPNN, adap i e ne wo k-based uzzy in e ence sys em (ANFIS), and SARIMA models,
and hus, used a di e en ial e olu ion me aheu is ic algo i hm o op imize he weigh s
o a hyb id model. Thei expe imen al case s udy showed ha hei p oposed me hod
pe o med be e han he h ee indi idual me hods and had highe accu acy.
In sum, se e al gene al conclusions can be d awn om he li e a u e using hyb id
models o explo e ime se ies o ecas ing. Fi s , in ecen yea s, he e has a g owing
numbe o s udies in es iga ing he impac o using combina ion heo y on o ecas ing
accu acy; hei objec i e is o enhance o ecas ing accu acy by combining di e en
Table 1 Li e a u e on se ies linea o nonlinea models o ime se ies o ecas ing
Au ho (s) Linea model Nonlinea model Yea Field
Ghasemi e al. (2016) ARIMA SVM 2016 Elec ici y p ice and load
o ecas ing
Ba ow (2016) SMA MLP 2016 In aday call a i als
o ecas ing
Ka is and Daskalaki (2015) FARIMA MLP 2015 In e ne a ic o ecas ing
Chaâbane (2014) FARIMA MLP 2014 Elec ici y p ice o ecas ing
Adhika i and Ag awal (2013) RWM MLP 2013 Financial ime se ies
o ecas ing ene gy
Wang and Meng (2012) ARIMA MLP 2012 Consump ion o ecas ing
Khashei e al. (2012) ARIMA PNNs 2012 Time se ies o ecas ing
Nou ani e al. (2011) SARIMAX MLP 2011 Rain all– uno p ocess
modeling
Wu and Chan (2011) ARIMA Time delay neu al
ne wo k (TDNN)
2011 Hou ly sola adia ion
o ecas ing
Aladag e al. (2009) ARIMA Elman’s ecu en
neu al ne wo ks
2009 Time se ies o ecas ing
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 3 o 24
models. The nume ical esul s o he e iewed pape s e idence ha he p edic i e cap-
abili y and accu acy o hyb id models a e be e han hose o single models. Mo eo e ,
hyb id models ha e ecen ly become a domina ed ool o ime se ies o ecas ing. Sec-
ond, schola s ha e in oduced se ies and pa allel combina ion me hodologies o con-
nec he componen s o hyb id models. Howe e , he ques ion o how o combine
single models, ha is, which combina ion yields mo e accu a e esul s, emains un-
answe ed. In o he wo ds, he li e a u e has neglec ed o compa e he wo ypes o
hyb id me hods o in oduce a mo e accu a e one and ocused on imp o ing o ecas-
ing accu acy by employing hyb id models a he han hei cons i uen s. Thi d, he li -
e a u e e iew e ealed ha among he linea and nonlinea models, ARIMA and
MLPs ha e a ac ed o e whelming a en ion and pe o m well when pa o hyb id
models gi en hei unique ea u es. ARIMA models a e one o he mos impo an
o ecas ing models ha ha e been success ully applied in modeling and o ecas ing.
The popula i y o he ARIMA model can be a ibu ed o i s s a is ical p ope ies and
he well-known Box–Jenkins (Box and Jenkins 1976) me hodology in he model-
building p ocess. The model assumes a linea co ela ion be ween he alues o a ime
se ies and hus, pe o ms well in linea modeling. MLP is he mos well-known a i icial
neu al ne wo k ha p ocesses nonlinea pa e ns in da a wi hou any assump ion and
does no equi e he de e mina ion o a model’s o m. MLPs a e lexible compu ing
amewo ks and uni e sal app oxima o s wi h a high deg ee o accu acy and can be
applied o a wide ange o o ecas ing p oblems. The key ad an age o he neu al
ne wo ks is hei lexible nonlinea modeling.
Gi en ha he li e a u e on ime se ies o ecas ing emains ambiguous on he choice o
combina ion s a egy, he co e objec i e o his s udy is o in oduce an e ec i e com-
bina ion me hodology and elucida e how indi idual models can be combined o imp o e
inancial ime se ies o ecas ing. Acco dingly, his s udy p esen s a comp ehensi e discus-
sion on se ies and pa allel combina ion me hods and hen, cons uc s a model using bo h
echniques o combine MLP as a nonlinea model and ARIMA as a linea model. Then,
using wo combina ion s a egies, ARIMA-MLP and MLP-ARIMA, he se ies and pa allel
hyb id models, comp ising simple a e age (SA), linea eg ession (LR) and gene ic
algo i hm (GA), a e compa ed wi h hei indi idual componen s. To e alua e he e ec-
i eness o he hyb id models and in oduce a mo e accu a e and eliable hyb id me hod,
wo benchma k da ase s, he closing o Shenzhen In eg a ed Index (SZII) and ha o
S anda d and Poo ’s 500 (S&P 500), a e selec ed o he o ecas ing and modeling.
The emainde o his pape is o ganized as ollows. Sec ion Me hods p esen s he basic
concep s and modeling p ocedu es o he ARIMA and MLP modes o ime se ies
o ecas ing. Sec ion Se ies combina ion me hod o ARIMA and MLP models and Pa allel
combina ion o ARIMA and MLP desc ibe he se ies and pa allel combina ion echniques
and he hyb id models cons uc ed using hese me hods. Sec ion Resul s and discussion
epo s he empi ical esul s o he hyb id se ies and pa allel models o a o ecas ing
benchma k da ase . Sec ion Compa ison o o ecas ing esul s compa es he pe o mance
o he models o he o ecas ing benchma k da ase . Sec ion Conclusions concludes.
Me hods
This sec ion in oduces he basic concep s and modeling p ocedu es o he ARIMA
and MLP models and se ies and he pa allel hyb id me hods o ime se ies o ecas ing.
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 4 o 24
ARIMA model
ARIMA is one o he mos widely used app oaches o p edic he u u e alue o ime
se ies by ex ac ing and modeling linea pa e ns in da a. The e o e, he classic model
is sui able o linea pa e ns. In ARIMA models, he u u e alue o a a iable is
assumed o be a linea unc ion o he pas alues and e o e ms.
y ¼u þφ1y −1þ…… þφpy −p−ε −θ1ε −1−……−θqε −qð1Þ
whe e (y
) is ac ual alue in ime and ε
is whi e noise, which is assumed o be
independen ly and iden ically dis ibu ed wi h a mean o ze o and cons an a iance o
σ
2
.pand qa e he in ege numbe s o au o eg essi e and mo ing a e age e ms in he
ARIMA model and φ
i
(i =1,2,...., p) and θ
i
(j =1,2,....,q) a e he model pa ame e s o be
es ima ed. The modeling p ocedu e o he ARIMA models, which is based on he
Box–Jenkins me hodology, comp ises h ee i e a i e s eps: model iden i ica ion, pa am-
e e es ima ion, and diagnos ic checking. In he iden i ica ion s ep, da a ans o ma ion
is o en equi ed o ende he ime se ies s a iona y, which is a necessa y condi ion
when building an ARIMA model o o ecas ing. A s a iona y ime se ies is cha ac e -
ized by a cons an mean and au oco ela ion s uc u e o e ime. When he obse ed
ime se ies p esen s a end and he e oscedas ici y, di e encing and powe ans o m-
a ion a e applied o he da a o emo e he end and s abilize he a iance be o e he
ARIMA model can be i ed. Once a en a i e model is iden i ied, he es ima ion o he
model pa ame e s is s aigh o wa d. The pa ame e s a e es ima ed such ha an o e all
measu e o e o s is minimized, which can be accomplished using a nonlinea
op imiza ion p ocedu e. The inal s ep is he diagnos ic checking o model adequacy,
which de e mines i he model assump ions abou e o s a
a e sa is ied.
Se e al diagnos ic s a is ics and esidual plo s can be used o examine he
goodness o i o a en a i ely adop ed model o he his o ical da a. I he model
is deemed inadequa e, a new en a i e model is iden i ied, which is also subjec ed
o pa ame e es ima ion and model e i ica ion. Diagnos ic in o ma ion can help
de e mine al e na i e model(s). This h ee-s ep model-building p ocess is ypically
epea ed se e al imes un il a sa is ac o y model is iden i ied. The inal model is
hen used o he p edic ion.
MLP model
Compu a ional in elligence sys ems, mo e speci ically, ANNs, which in ac , a e a ee
dynamics model, a e being widely used o he app oxima ion o unc ions and
o ecas ing. In he case o eal-wo ld p oblems, neu al ne wo ks a e an e ec i e ool o
ecognize nonlinea pa e ns. ANNs a e uni e sal app oxima o s ha app oxima e a
la ge class o unc ions wi h a high deg ee o accu acy, which is a c ucial ad an age
o e o he classes o nonlinea models (Zhang e al. 1998). Thei powe is de i ed om
he pa allel p ocessing o in o ma ion om he da a and no p io assump ion is e-
qui ed in he model-building p ocess. Ins ead, he ne wo k model is la gely de e mined
by he da a cha ac e is ics. MLPs o single hidden laye eed- o wa d neu al ne wo ks
a e key and commonly used model o ms o ANNs o ime se ies modeling and
o ecas ing. The model is cha ac e ized by a ne wo k o h ee laye s o simple p ocess-
ing uni s connec ed by acyclic links (Fig. 1). The ela ionship be ween he ou pu (y
)
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 5 o 24
and (y
−1
,…,y
−p
) inpu s has he ollowing ma hema ical ep esen a ion (Khashei and
Bija i 2010):
y ¼w0þX
q
j¼1
wj⋅gw
0;jþX
p
i¼1
wi;j⋅y −i
!
þε ;ð2Þ
whe e w
i, j
(i=0,1,2,..., p, j =1,2,...,q) and w
j
(j=0,1,2,...,q) a e model pa ame e s o en
e med connec ion weigh s, pis he numbe o inpu nodes, and qis he numbe o
hidden nodes. The ac i a ion unc ions ake se e al o ms.
The ype o ac i a ion unc ion is indica ed by he condi ion o he neu on wi hin he
ne wo k. In a majo i y o cases, inpu laye neu ons do no ha e an ac i a ion unc ion
because hei ole is o ans e inpu s o he hidden laye . The mos widely used
ac i a ion unc ion o he ou pu laye is he linea unc ion because a non-linea one
may dis o he p edica ed ou pu . The logis ic unc ion is o en used as a hidden laye
ans e unc ion, as shown in Eq. (3). O he ac i a ion unc ions can also be used such
as linea and quad a ic unc ions, each wi h a a ie y o modeling applica ions.
Sig xðÞ¼ 1
1þexp −xðÞ
:ð3Þ
Fig. 1 MLP neu al ne wo k a chi ec u e
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 6 o 24
Thus, he ANN model in Eq. (2) pe o ms a nonlinea unc ional mapping om he
pas obse a ions o he u u e alue y, ha is,
y ¼ ðy −1;…;y −p;wÞþε ;ð4Þ
whe e wis a ec o o all pa ame e s and (.) is a unc ion de e mined by he ne -
wo k s uc u e and connec ion weigh s. Thus, he MLP is equi alen o a nonlinea
au o eg essi e model. The simple ne wo k in Eq. (2) is unexpec edly powe ul, ha
is, i can app oxima e he a bi a y unc ion as a numbe o hidden nodes when q
is su icien ly la ge. In p ac ice, a simple ne wo k s uc u e wi h a small numbe o
hidden nodes o en wo ks well in ou -o -sample o ecas ing, possibly because o
he o e - i ing e ec ypically ound in he MLP modeling p ocess. An o e - i ed
model has a good i o he sample used o model building bu poo
gene alizabili y o ou -o -sample da a.
The choice o qis da a dependen and he e is no sys ema ic ule in deciding
his pa ame e . In addi ion o choosing an app op ia e numbe o hidden nodes,
selec ing he numbe o lagged obse a ions, p, and dimensions o he inpu
ec o is an impo an ask in he ANN modeling o a ime se ies. This is,
pe haps, he mos impo an pa ame e o be es ima ed in an ANN model
because i plays a majo ole in de e mining he (nonlinea ) au oco ela ion
s uc u e o he ime se ies.
Se ies combina ion me hod o ARIMA and MLP models
In he se ies linea o nonlinea combina ion models, a ime se ies is di ided in o a
linea and nonlinea pa , as ollows:
y ¼X
n
¼1
L þX
n
¼1
N ð5Þ
whe e L
and N
deno e he linea and nonlinea pa s es ima ed om he da a.
Then, hese wo componen s a e sequen ially p ocessed by ARIMA and MLP
models. Thus, in he i s s age o his me hod, he ARIMA o MLP model is
selec ed o iden i y linea o nonlinea pa e ns in he o iginal da a. Then, o
disco e he emaining pa e ns ha a e no cap u ed by he i s model, he
ou pu ob ained in he i s s age is used as an inpu o he second model. The
basic concep is ha one model is insu icien o cap u e all ela ionships in he
da a. Mo eo e , ully iden i ying and modeling he da a cha ac e is ics in he eal
ime se ies is di icul and some imes, e en impossible. Thus, using an indi idual
model such as he ARIMA (MLP) model, undoub edly, e eals nonlinea (linea )
pa e ns ha a e no comple ely ecognized. Consequen ly, he MLP (ARIMA)
model is employed in he second s age o cap u e he emaining nonlinea
(linea ) pa e ns. A summa ion o he ou pu s ob ained om he wo s ages is
conside ed he inal combined o ecas . On he basis o he sequence o model
selec ion, wo hyb id models (ARIMA-MLP and MLP-ARIMA) a e p esen ed in
he nex sec ion.
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 7 o 24
ARIMA-MLP model
In line wi h he se ies modeling p ocedu e, in he i s s age o he ARIMA-MLP
model, he ARIMA is applied o model he linea componen . Le e
deno e he esidual
o he ARIMA model a ime :
e ¼y −
^
L ð6Þ
whe e
^
L is he o ecas ing alue o ime om he ARIMA model based on o iginal
da a. In he second s age, he esiduals o he i s s age a e used as inpu da a o he
MLP model, allowing o he iden i ica ion o nonlinea ela ionships. Wi h ninpu
nodes, he MLP model o he esiduals will be.
e ¼ ðe −1;e −2;…;e −nÞþε ⇒c
N′
¼^
e ¼ ðe −1;e −2;…;e −nÞð7Þ
whe e is a nonlinea unc ion de e mined by he MLP, N′
is he o ecas ing alue o
ime in he MLP model based on esidual da a, and e
is he andom e o . The ame-
wo k o he ARIMA-MLP model is displayed in Fig. 2a.
No e ha i model is inapp op ia e, he e o e m is no necessa ily andom;
he e o e, he co ec iden i ica ion is c i ical. In his way, he combined o ecas will
be as ollows:
^
y ¼
^
L þc
N
′
ð8Þ
MLP-ARIMA model
Simila o he ARIMA-MLP model, he MLP-ARIMA model has wo main s ages. In
he i s s age, he MLP model is used o model he nonlinea pa o he ime se ies.
Le e′
deno e he esidual o he MLP model a ime . Then,
(a)
(b)
Fig. 2 F amewo k o (a) ARIMA-MLP and (b) MLP-ARIMA model
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 8 o 24
Fig. 8 Es ima ed alues o GA-based hyb id model o SZII
Fig. 7 Es ima ed alues o SA-based hyb id model o SZII
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 15 o 24
S anda d and Poo ’s 500 da ase
S anda d and Poo ’s 500 (S&P 500) da ase includes 2349 daily closing s ock p ices
om Oc obe 1998 o Feb ua y 2008 (Zhang and Wu 2009). The S&P 500 da ase is
plo ed in Fig. 10. Acco ding o p e ious s udies, he S&P da ase is di ided in o
aining and es da ase s. The i s obse a ions (abou 80% o he sample) a e used as
he aining sample o o mula e he models and he las 470 obse a ions a e applied
as a es sample o e alua e he pe o mance o he cons uc ed models.
ARIMA-MLP se ies hyb id model
S age I - (Linea modeling): Simila o he linea modeling phase, ARIMA(1,0,0) is
designed and he esiduals o his s ep a e used in he nex s ep.
Table 2 Pe o mance o models o SZII using ain and es da ase s
Model T ain Tes
MAE MAE MAPE RMSE MAE MSE MAPE RMSE
ARIMA-MLP 212.39 91,955 7.22% 303.24 1082.88 1,915,716 9.52% 1384.09
MLP-ARIMA 210.91 86,221 7.23% 293.63 1064.91 1,915,422 10.16% 1383.98
SAHM 217.78 95,399 7.12% 308.86 1123.87 1,997,323 9.91% 1388.69
GAHM 215.38 94,577 7.08% 307.53 1102.75 1,969,593 9.79% 1403.42
LRHM 215.54 94,573 7.09% 307.52 1074.87 1,928,479 9.64% 1388.69
ARIMA 224.45 99,255 7.30% 315.04 1166.17 2,221,776 10.26% 1490.56
MLP 215.38 94,577 7.30% 307.53 1102.33 1,974,479 9.80% 1405.16
Fig. 9 Es ima ed alues o LR-based hyb id model o SZII
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 16 o 24
S age II - (Nonlinea modeling): In his s age, he esiduals o he p e ious s ep a e
used as inpu o he MLP model and a ne wo k wi h h ee inpu , i e hidden, and one
ou pu neu on is i ed o ex ac he emaining nonlinea s uc u es.
S age III - (Combina ion): He e, he o ecas ed alues o p e ious wo s ages a e
combined o gene a e he inal combined o ecas . The es ima ed alues o he
ARIMA-MLP model agains he ac ual alues o all da a a e plo ed in Fig. 11.
MLP-ARIMA se ies hyb id model
S age I - (Nonlinea modeling): In he nonlinea modeling phase, a ne wo k wi h h ee
inpu , h ee hidden, and one ou pu neu ons is designed o cap u e he nonlinea
ela ionships in he ime se ies gene a ed and he gene a ed esiduals a e used in he
nex s ep.
S age II: (Linea modeling): In his s ep, an ARIMA (3, 0, 3) model is i o p ocess he
linea s uc u es ha a e no modeled by he MLP model.
S age III: (Combina ion): In he inal s ep, he o ecas ed alues om s ages I and II
a e combined. The es ima ed alues o he MLP-ARIMA model agains he ac ual
alues o all da a a e plo ed in Fig. 12.
Pa allel hyb id models
S age I - (Linea and nonlinea modeling): Simila o he p e ious sec ion, o cap u e
he linea and nonlinea pa e ns in he da a o he S&P ime se ies, he ARIMA (1, 0, 0)
and MLP models wi h h ee inpu , h ee hidden, and one ou pu neu on a e designed.
Fig. 10 Daily S&P 500 s ock closing p ices, Oc obe 1998–Feb ua y 2008
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 17 o 24
Fig. 12 Es ima ed alues o MLP-ARIMA model o S&P500
Fig. 11 Es ima ed alues o ARIMA-MLP model o S&P 500
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 18 o 24
Fig. 13 Es ima ed alues o SA-based hyb id model o S&P 500
Fig. 14 Es ima ed alues o GA-based hyb id model o S&P 500
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 19 o 24
S age II - (Ini ializing weigh s): In his s a e, he op imum weigh s a e de i ed by
applying he LR, GA, and SA weigh ing me hods. No e ha he OLS app oach and
GA a e designed using E iwes and MATLAB so wa e.
S age III - (Combina ion): Acco ding o he modeling p ocedu e o he pa allel hyb id
models, he combined o ecas is made using he alues ob ained om he p e ious wo
s ages. The es ima ed alues o he hyb id models based on pa allel SA, GA, and LR
agains he ac ual alues o all da a a e plo ed in Figs. 13,14,15, espec i ely.
Table 3 summa izes he pe o mance o he se ies and pa allel hyb id models in
p edic ing he S&P 500 s ock p ice using ain and es da ase s. The able shows ha
he se ies models, ARIMA-MLP and ARIMA-MLP, a e no compa able wi h each o he
and epo s signi ican ly high p edic ion pe o mance when he se ies hyb id
me hodology is used ins ead o he pa allel me hods and hei base models.
Fig. 15 Es ima ed alues o LR-based hyb id model o S&P 500
Table 3 Pe o mance o models o S&P 500 using ain and es da ase s
Model T ain Tes
MAE MSE MAPE RMSE MAE MSE MAPE RMSE
ARIMA-MLP 9.94 176.76 0.01% 13.30 9.02 159.33 0.64% 12.62
MLP-ARIMA 9.96 180.17 0.01% 13.42 9.20 157.69 0.65% 12.56
SAHM 9.98 180.69 0.01% 13.44 9.22 164.48 0.65% 12.83
GAHM 9.97 180.31 0.65% 13.42 9.23 159.33 0.85 12.62
LSHM 9.97 180.28 0.01% 13.43 8.98 152.55 0.63% 12.35
ARIMA 10.02 182.21 086% 13.50 9.36 173.85 0.66% 13.19
MLP 9.97 180.31 0.01% 13.43 9.24 159.34 0.65% 12.62
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 20 o 24
Compa ison o o ecas ing esul s
This sec ion compa es he p edic i e capabili ies o he hyb id models cons uc ed by
applying he se ies and pa allel combina ion me hods wi h ei he o hei componen s,
MLP, and ARIMA, using he wo abo emen ioned da ase s. The compa a i e analysis is
conduc ed om wo iewpoin s: compa ison o se ies and pa allel hyb id models and
analysis o a e age pe cen age imp o emen in he se ies and pa allel hyb id models in
compa ison wi h hei componen s. Two pe o mance indica o s, MAE and MSE,
a e employed o compa e he o ecas ing pe o mance o he hyb id models and
hei componen s.
In he i s s ep, he o e all pe o mance o he se ies and pa allel hyb id models is
compa ed. Tables 4 and 5 p esen he o e all pe o mance o he hyb id models and
hei componen s o he SZII and S&P 500 da ase s. The compa ison e eals ha he
a e age o ecas ing e o o MAE and MSE in he ain and es da ase s is lowe in
he ARIMA-MLP and MLP-ARIMA hyb id models cons uc ed using he se ies
combina ion echnique han hose in he pa allel hyb id models. Fo example, in he
SZII da ase , in MAE and MSE e ms, he o ecas ing esul s o he se ies models using
es da ase imp o ed by 2.41% and 2.51% compa ed o hose o he pa allel hyb id
models. Tables 4 and 5 compa e he pe o mance and accu acy o he se ies and
pa allel hyb id me hods. In he second s ep, he pe o mance o he hyb id models is
compa ed wi h hose o hei base models. In o he wo ds, he a e age pe cen age
imp o emen o he se ies and pa allel hyb id models is compa ed wi h ha o hei
base models. The esul s show ha applying all i e hyb id models, on a e age, im-
p o es he o ecas ing accu acy o e a leas one model o he ARIMA and MLP
neu al ne wo k models. This con i ms he hypo hesis ha indi idual models do no
cap u e all ela ionships in he da a and combining he wo models can be e ec i e in
o e coming hei limi a ions and imp o ing o ecas ing accu acy.
Tables 6 and 7 p esen he a e age imp o emen pe cen age o he se ies and pa allel
hyb id models o he SZII and S&P 500 da ase s compa ed o he ARIMA and MLP
Table 4 O e all pe o mance o se ies and pa allel models o SZII
Model T ain Tes
MAE MSE MAE MSE
Se ies models 211.65 89,088.05 1073.89 1,915,569.00
Pa allel models 216.23 94,849.02 1100.49 1,965,131.66
ARIMA 224.45 99,255.51 1166.17 2,221,776.70
MLP 215.38 94,577.00 1102.33 1,974,479.94
Lowe o ecas ing e o a e in bold
Table 5 O e all pe o mance o se ies and pa allel models o S&P 500
Model T ain Tes
MAE MSE MAE MSE
Se ies models 9.95 178.46 9.11 158.51
Pa allel models 9.97 180.42 9.14 158.78
ARIMA 10.02 182.21 9.36 173.85
MLP 9.97 180.31 9.24 159.34
Lowe o ecas ing e o a e in bold
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 21 o 24
models. The esul s sugges ha , on a e age, he se ies hyb id models ha e a highe
imp o ing impac on he ARIMA and MLP models han he pa allel hyb id models.
Fo example, when using he S&P 500 da ase , he se ies hyb id models imp o e he
MLP models in e ms o MSE by 1.02% in he case o ain da a, while his imp o e-
men is −0.21 o he pa allel models. Acco ding o he analy ical esul s, he accu acies
o he se ies hyb id models a e be e han hose o he pa allel hyb id models in bo h
o e all pe o mance and a e age imp o emen pe cen age o e he base models. F om
he abo e compa a i e analyses, he models can be anked as ollows: (i) se ies hyb id
models (ii) pa allel hyb id models, and (ii) indi idual models.
Conclusions
Fo ecas ing eal-wo ld ime se ies, pa icula ly inancial ime se ies, is a c i ical ask
ha has ecen ly ecei ed o e whelming a en ion. Gi en he impo ance o accu a e
o ecas ing, se e al ela ed me hods ha e been p oposed in he li e a u e. In addi ion o
single me hods, s udies ha e combined di e en me hods o gene a e mo e accu a e e-
sul s and con i med ha combining di e en models enhances o ecas ing accu acy
and accoun s o he unique ea u es o indi idual models. Al hough nume ous s udies
ha e used se ies o pa allel me hods o cons uc hyb id models and con i m ha com-
bining di e en models educes o ecas ing e o and o e mo e accu a e esul s, hey
emain ague on he p ecise combina ion ha p oduces a mo e accu a e hyb id model.
Thus, his s udy p oposed a mo e e icien echnique o o ecas inancial ime se ies
and hen, conduc ed a comp ehensi e compa ison o he p edic i e capabili ies o he
se ies and pa allel combina ion echniques ha we e combined wi h linea and nonlin-
ea models, such as ARIMA and MLP, along wi h hei indi idual componen s. Fi s ,
he se ies and pa allel hyb id models we e compa ed, ollowed by a compa ison o he
hyb id models’a e age pe cen age imp o emen wi h hose o hei base models. The
empi ical esul s o he wo benchma k da ase s, SZII and S&P 500, indica ed ha all
hyb id models cons uc ed using he wo combina ion me hods gene a ed supe io
esul s han a leas one o hei indi idual componen s. The esul s also show ha he
se ies me hod gene a e mo e accu a e hyb id models and has a highe imp o emen
Table 6 A e age imp o emen in se ies and pa allel models o e ARIMA and MLP models o SZII
Hyb id model ARIMA MLP
T ain (%) Tes (%) T ain (%) Tes (%)
MAE MSE MAE MSE MAE MSE MAE MSE
Se ies models 5.70 10.24 7.91 13.78 1.72 5.80 2.57 2.98
Pa allel models 3.65 4.43 5.64 11.55 −1.11 −0.40 0.27 0.47
Lowe o ecas ing e o a e in bold
Table 7 A e age imp o emen in se ies and pa allel models o e ARIMA and MLP models o
S&P 500
Hyb id model ARIMA MLP
T ain (%) Tes (%) T ain (%) Tes (%)
MAE MSE MAE MSE MAE MSE MAE MSE
Se ies models 0.69 2.05 2.66 8.82 0.3 1.02 1.40 1.25
Pa allel models 0.49 0.63 1.45 6.37 0.00 −0.21 1.04 0.52
Lowe o ecas ing e o a e in bold
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 22 o 24
pe cen age han he pa allel me hod. The e o e, he se ies combina ion me hod can be
conside ed an e icien al e na i e o cons uc mo e accu a e hyb id models in bo h
analy ical app oaches o o ecas ing inancial ime se ies.
Fu u e wo ks should conside implemen ing he se ies and pa allel hyb id
me hodologies o de elop an app oach wi h h ee o mo e indi idual models and
acco dingly, compa e and analyze he ob ained esul s. Resea che s can also examine
o he s a is ical and in elligen models, such as GARCH and SVM models, o cons uc
se ies and pa allel hyb id models o o ecas inancial ime se ies.
Acknowledgemen s
The au ho s exp ess hei g a i ude o D . Fa imah Mokha ab Ra iei, associa e p o esso o indus ial enginee ing a he
Ta bia Moda es Uni e si y o Teh an, and D . Mehdi Bija i, p o esso o indus ial enginee ing a Is ahan Uni e si y o
Technology, o hei insigh ul and cons uc i e commen s, which ha e helped conside ably imp o e his pape .
Funding
The au ho s ha e no unding o epo .
A ailabili y o da a and ma e ials
No applicable.
Au ho s’con ibu ions
All au ho s ha e equally con ibu ed o his wo k and app o e o his submi s.
Compe ing in e es s
The au ho s decla e ha hey ha e no compe ing in e es s.
Publishe ’sNo e
Sp inge Na u e emains neu al wi h ega d o ju isdic ional claims in published maps and ins i u ional a ilia ions.
Recei ed: 15 Ma ch 2017 Accep ed: 20 Oc obe 2017
Re e ences
Adhika i R, Ag awal RK (2013) A combina ion o a i icial neu al ne wo k and andom walk models o inancial ime
se ies o ecas ing. J Neu al Compu Appl 22:1–9
Aghay Kaboli SH, Sel a aj j, Rahim NA (2016a) Long- e m elec ic ene gy consump ion o ecas ing ia a i icial
coope a i e sea ch algo i hm. J Ene gy 115:857–871
Aghay Kaboli SH, Fallahpou A, Kazem N, Sel a aj J, Rahim NA (2016b) An exp ession-d i en app oach o long- e m
elec ic powe consump ion o ecas ing. J Da a Min Knowl Disco 1(1):16–28
Aghay Kaboli SH, Sel a aj j, Rahim NA (2017a) Rain- all op imiza ion algo i hm: a popula ion based algo i hm o sol ing
cons ained op imiza ion p oblems. J Compu Sci 19:31–42
Aghay Kaboli SH, Fallahpou A, Sel a aj J, Rahim NA (2017b) Long- e m elec ical ene gy consump ion o mula ing and
o ecas ing ia op imized gene exp ession p og amming. J Ene gy 126:144–164
Aladag CH, Eg ioglu E, Kadila C (2009) Fo ecas ing nonlinea ime se ies wi h a hyb id me hodology. J Appl Ma h Le
22:1467–1470
Ba ow DK (2016) Fo ecas ing in aday call a i als using he seasonal mo ing a e age me hod. J Bus Res 69:6088–6096
Ba es JM, G ange WJ (1969) The combina ion o o ecas s. J Ope Res 20:451–468
Box P, Jenkins G (1976) Time se ies analysis: o ecas ing and con ol. Holden-day Inc, San F ancisco, CA
Bunn D (1989) Fo ecas ing wi h mo e han one model. J Fo ecas ing 6:161–166
Chaâbane N (2014) A hyb id ARFIMA and neu al ne wo k model o elec ici y p ice p edic ion. J Elec Pow Ene gy Sys
55:187–194
Chen KY, Wang CH (2007) A hyb id SARIMA and suppo ec o machines in o ecas ing he p oduc ion alues o he
machine y indus y in Taiwan. J Expe Sys Appl 32:254–264
De Menezes L, Bunn D (1993) Diagnosi ic acking and model speci ica ion in combined o ecas s o U.K. in la ion.
J Fo ecas ing 12:559–572
Ghasemi A, Shayeghi H, Mo adzadeh M, Nooshya M (2016) A no el hyb id algo i hm o elec ici y p ice and load
o ecas ing in sma g ids wi h demand-side managemen . J Appl Ene gy 177:40–59
G ange CWJ, Ramana han R (1984) Imp o ed me hods o combining o ecas s. J Fo ecas ing 3:197–204
Kaboli S H A; Fallahpou A, Kazemi N, Sel a aj J, Rahim N.A (2016) Elec ic ene gy consump ion o ecas ing ia
exp ession-d i en app oach. 4 h IET Clean Ene gy and Technology Con e ence
Ka is C, Daskalaki S (2015) Compa ing o ecas ing app oaches o in e ne a ic. J Expe Sys Appl 42:8172–8183
Khashei M, Bija i M (2010) An a i icial neu al ne wo k (p, d, q) model o ime se ies o ecas ing. J Exp Sys Appl
37:479–489
Khashei M, Bija i M, Raissi A dali GA (2012) Hyb idiza ion o au o eg essi e in eg a ed mo ing a e age (ARIMA) wi h
p obabilis ic neu al ne wo ks (PNNs). J Compu Indus Eng 63:37–45
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 23 o 24
Mak idakis S, Ande sen A, Ca bone R, Fildes R, Hibon M, Lewandowski R, W on J, Pa zen E, Winkle R (1982) The
accu acy o ex apola ion (TimeSe ies) me hods: esul s o a o ecas ing compe i ion. J Fo ecas ing 1:111–153
Modi i-Delshad M, Aghay Kaboli SH, Taslimi-Renani E, Abd Rahim N (2016) Back acking sea ch algo i hm o sol ing
economic dispa ch p oblems wi h al e-poin e ec s and mul iple uel op ions. J Ene gy 116:637–649
Nou ani V, Kisi O, Komasi M (2011) Two hyb id a i icial in elligence app oaches o modeling ain all– uno p ocess.
J Hyd ol 402:41–59
Pai PF, Lin CS (2005) A hyb id ARIMA and suppo ec o machines model in s ock p ice o ecas ing. J Omega
33:497–505
Pao TH (2009) Fo ecas ing ene gy consump ion in Taiwan using hyb id nonlinea models. J Ene gy 34:1438–1446
Ra he AM, Aga wal A, Sas y VN (2015) Recu en neu al ne wo k and a hyb id model o p edic ion o s ock e u ns.
J Exp Sys Appl 42:3234–3241
Wang X, Meng M (2012) A hyb id neu al ne wo k and ARIMA model o ene gy consump ion o ecas ing. J Compu
Secu 7(5):1184–1190
Wang J-J, Wang J-Z, Zhang Z-G, Guo S-P (2012) S ock index o ecas ing based on a hyb id model. J Omega 40:758–766
Wedding DK II, Cios KJ (1996) Time se ies o ecas ing by combining RBF ne wo ks, ce ain y ac o s, and he
box-Jenkins model. J Neu o-Oncol 10:149–168
Wu J, Chan CK (2011) P edic ion o hou ly sola adia ion using a no el hyb id model o ARMA and TDNN.
J Sola Ene gy 85:808–817
Yang Y, Chen Y, Wang Y, Li C, Li L (2016) Modelling a combined me hod based on ANFIS and neu al ne wo k
imp o ed by DE algo i hm: a case s udy o sho - e m elec ici y demand o ecas ing. J Appl So Compu
49:663–675
Zeng D, Xu J, Gu J, Liu L, Xu G (2008) Sho e m a ic low p edic ion using hyb id ARIMA and ANN models.
In: Powe Elec onics and In elligen T anspo a ion Sys em, pp 621–625
Zhang GP (2003) Time se ies o ecas ing using a hyb id ARIMA and neu al ne wo k model. J Neu o-Oncol 50:159–175
Zhang Y, Wu L (2009) S ock ma ke p edic ion o S&P 500 ia combina ion o imp o ed BCO app oach and BP neu al
ne wo k. J Exp Sys Appl 36:8849–8854
Zhang G, Pa uwo BE, Hu MY (1998) Fo ecas ing wi h a i icial neu al ne wo ks: he s a e o he a . J In J Fo ecas
14:35–62
Zhou ZHJ, Hu C-H (2008) An e ec i e hyb id app oach based on g ey and ARMA o o ecas ing gy o d i . J Chaos
Soli ons F ac als 35:525–529
Khashei and Haji ahimi Financial Inno a ion (2017) 3:24 Page 24 o 24
