Mu h, Manuel; Lingen elde , Michael; Nu e , Ge d
A icle — Published Ve sion
The applica ion o machine lea ning o demand
p edic ion unde mac oeconomic ola ili y: a sys ema ic
li e a u e e iew
Managemen Re iew Qua e ly
P o ided in Coope a ion wi h:
Sp inge Na u e
Sugges ed Ci a ion: Mu h, Manuel; Lingen elde , Michael; Nu e , Ge d (2024) : The applica ion o
machine lea ning o demand p edic ion unde mac oeconomic ola ili y: a sys ema ic li e a u e
e iew, Managemen Re iew Qua e ly, ISSN 2198-1639, Sp inge In e na ional Publishing, Cham,
Vol. 75, Iss. 3, pp. 2759-2802,
h ps://doi.o g/10.1007/s11301-024-00447-8
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/330566
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 ps://c ea i ecommons.o g/licenses/by/4.0/
Vol.:(0123456789)
Managemen Re iew Qua e ly (2025) 75:2759–2802
h ps://doi.o g/10.1007/s11301-024-00447-8
The applica ion o machine lea ning o demand p edic ion
unde mac oeconomic ola ili y: asys ema ic li e a u e
e iew
ManuelMu h1 · MichaelLingen elde 1· Ge dNu e 2
Recei ed: 20 No embe 2023 / Accep ed: 23 May 2024 / Published online: 6 June 2024
© The Au ho (s) 2024
Abs ac
In a con empo a y con ex cha ac e ised by shi s in mac oeconomic condi ions and
global unce ain y, p edic ing he u u e beha iou o demande s is c i ical o man-
agemen science disciplines such as ma ke ing. Despi e he ecognised po en ial o
Machine Lea ning, he e is a lack o e iews o he li e a u e on he applica ion o
Machine Lea ning in p edic ing demande s’ beha iou in a ola ile en i onmen .
To ill his gap, he ollowing sys ema ic li e a u e e iew p o ides an in e discipli-
na y o e iew o he esea ch ques ion: “How can Machine Lea ning be e ec i ely
applied o p edic demand pa e ns unde mac oeconomic ola ili y?” Following
a igo ous e iew p o ocol, a li e a u e sample o s udies (n = 64) is iden i ied and
analysed based on a hyb id me hodological app oach. The indings o his sys em-
a ic li e a u e e iew yield no el insigh s in o he concep ual s uc u e o he ield,
ecen publica ion ends, geog aphic cen es o scien i ic ac i i y, as well as leading
sou ces. The esea ch also discusses whe he and in which ways Machine Lea ning
can be used o demand p edic ion unde dynamic ma ke condi ions. The e iew
ou lines a ious implemen a ion s a egies, such as he in eg a ion o o wa d-look-
ing da a wi h economic indica o s, demand modelling using he Coe icien o Va i-
a ion, o he applica ion o combined algo i hms and speci ic A i icial Neu al Ne -
wo ks o accu a e demand p edic ions.
Keywo ds Machine lea ning· Mac oeconomic ola ili y· Demand o ecas ing·
Ma ke ing p edic ions· Sys ema ic li e a u e e iew
JEL Classi ica ion C53· E32· C45· M31
* Manuel Mu h
[email p o ec ed]
1 School o Business andEconomics, Philipps-Uni e si ä Ma bu g, Uni e si ä ss . 24,
35037Ma bu g, Ge many
2 ESB Business School, Reu lingen Uni e si y, Al ebu gs . 150, 72762Reu lingen, Ge many
2760
M.Mu h e al.
1 In oduc ion
In ligh o geopoli ical ins abili ies in Eas e n Eu ope and he Middle Eas , changes
in in la ion and in e es a es, and dis up ions in global supply chains (Dö ne e al.
2023; Eu opean Cen al Bank 2023), a ious business unc ions ace a complex ask
in accu a ely p edic ing he beha iou o demande s. The economic ci cums ances
in luencing manage ial ope a ions ha e hence unde gone p o ound changes and
many o he exis ing p edic ion app oaches ely on subs an ially di e en ci cum-
s ances han hose cu en ly p e ailing (Du s e al. 2022; Du ugbo and Al-Balushi
2022). As a esul , “con empo a y o ganiza ions ace en i onmen s wi h unp ec-
eden ed le els o ola ili y, unce ain y, complexi y, and ambigui y” (T oise e al.
2022), p esen ing new challenges o p edic ing he u u e beha iou o demande s.
In scien i ic li e a u e, se e al au ho s emphasise he gene al po en ial o Machine
Lea ning (ML) o p edic i e analy ical asks in impo an business a eas like ma -
ke ing (Huang and Rus 2021; Ma and Sun 2020; Ve ma e al. 2021). Howe e ,
he e is an exis ing knowledge gap in e ms o p edic ing u u e demand wi h ML
in ola ile en i onmen s (Ghoddusi e al. 2019), lea ing a signi ican need o u -
he esea ch. The lack o comp ehensi e li e a u e e iews on his pa icula ques-
ion is subs an ia ed by p elimina y sea ches in he Web o Science-da abase ac oss
12,000 a ailable jou nals and sc eening he 205 que y esul s ha include he i le
wo ds “sys ema icli e a u e e iew” and “machine lea ning”, along wi h addi ional
in es iga ions in u he elec onic esou ces. While some exis ing li e a u e e iews
add ess he o e a ching in e sec ion o ML and applied managemen disciplines,
he e is a endency o ocus mo e on gene al ques ions such as p ima y applica ion
ca ego ies o global end de elopmen s (see u he Mus ak e al. 2021; Ve ma e al.
2021; Vlačić e al. 2021). Al hough he e a e al eady e iew pape s ha add ess he
applica ion o ML and A i icial In elligence, each adop ing di e en speci ic pe -
spec i es (e.g., Kaushal e al. 2023; Keding 2021), none o he a ailable e iews
aligns wi h he hema ic ocus o his sys ema ic li e a u e e iew.
This backg ound p o ides he unde lying ounda ion o de ining he esea ch
ques ion o his sys ema ic li e a u e e iew. The gene al scien i ic p oblem is s uc-
u ed acco ding o he “CIMO” logic (Con ex , In e en ion, Mechanism, Ou come),
o mula ed by Denye e al. (2008) and ecen ly add essed by Kucke z and Block
(2021). As a esul , he esea ch ques ion o his pape is de ined as: “How can
ML (I) be e ec i ely applied (M) o p edic demand pa e ns (O) unde mac oeco-
nomic ola ili y (C)?”. The aim is o collec and syn hesise s a e-o - he-a academic
knowledge su ounding his ques ion while adop ing an applica ion pe spec i e om
ma ke ing science.
In line wi h he ecommenda ion o Linnenluecke e al. (2019), he aim o his
wo k is “ o conside in e disciplina y con ibu ions” o ensu e an in eg a i e o e -
iew co e ing he ele an sou ces o in o ma ion. To os e such an in e discipli-
na y synopsis, me hodological concep s om compu e science a e adop ed wi h a
p ac ical business pe spec i e o ma ke ing, while main aining an inclusi e app oach
owa ds o he managemen disciplines such as economics. The esea ch i sel is
con ined o a p ima ily analy ical iewpoin wi h an applica ion-o ien ed ocus. In
2761
The applica ion o machine lea ning o demand p edic ion…
s ic acco dance o he esea ch ques ion, he aim o his e iew pape is o add ess
he ask o demand p edic ion, and he e o e no ocus is placed on explana o y o
desc ip i e asks (see u he Thommen e al. 2017). Consequen ly, in en ionally
ou side he scope o his esea ch a e a p ecise modelling o indi idual decision-
making mechanisms, an explana ion o consume beha iou o an unde s anding o
he unde lying mo i a ions and p e e ences o demande s. Mo eo e , his e iew
pape ollows he gene al unde s anding ha ML applica ions a e inhe en ly obse -
a ion-d i en (see u he Ghoddusi e al. 2019; Xie 2020). The e o e, he demand
p edic ions discussed he e a e based on an exis ing da ase and, consequen ly, no
guided by gene al economic heo ies o he unc ional ela ionships o ma ke pa -
icipan s ha a e no e lec ed in he unde lying da a.
The s uc u e o his e iew pape is buil on he me hodology by Xiao and Wa son
(2019) o sys ema ic li e a u e e iews. The co e s ages o hei me hodology p o ide
gene al guidance o he s uc u ing o his e iew pape and consis o planning he
e iew, conduc ing he e iew, and epo ing on he e iew. Based on his, he e iew
pape is s uc u ed as ollows: A e he in oduc ion (1), a b ie heo e ical and e -
minological ounda ion is p o ided (2). Subsequen ly, he esea ch me hodology o
he li e a u e e iew is p esen ed (3), illus a ing he planning o he e iew. The ol-
lowing sec ion (4) hen deals wi h he implemen a ion o he e iew and in ol es an
analysis o he li e a u e sample. In his way, he s eps o sea ching and selec ing li e -
a u e, as well as quali y e alua ion and da a ex ac ion, a e demons a ed. A e wa ds,
an analysis o he a icle cha ac e is ics (4.1), an analysis o he abs ac s (4.2), and an
analysis o he ull ex s (4.3) a e p o ided o he inal li e a u e sample— ollowed
by he gene al limi a ions (4.4) o he s udy. The inal sec ion (5) p o ides a conclu-
si e epo on he key componen s and o e s a comp ehensi e conclusion.
2 Theo e ical backg ound
2.1 Machine lea ning
Lanquillon (2019) dis inguishes ML by enabling he gene a ion o a model om a
da ase by a lea ning p ocedu e ins ead o by an explici p og amming ins uc ion. A
mo e gene al de ini ion is p o ided by Mi chell (1997), who desc ibes ML as being
able “ o lea n om expe ience E wi h espec o some class o asks T and pe -
o mance measu e P, i i s pe o mance a asks in T, as measu ed by P, imp o es
wi h expe ience E.” O e he las decades, he unde s anding o ML as a sub- ield
o A i icial In elligence (AI) has become gene ally accep ed (Ma and Sun 2020;
Shaikh e al. 2022; Ve ma e al. 2021). F om an econome ics pe spec i e, Ghod-
dusi e al. (2019) obse e ha economis s adi ionally end o ocus on heo e ically
guided modelling in ol ing s a is ical analysis o indi idual explo a o y a iables,
whe eas he p ima y ocus o ML is on gene a ing a p edic ion using a ailable da a
inpu s. In iew o his, ML is cha ac e ised as obse a ion-d i en modelling, which
is gene ally di e en om he way pa ame ic models a e gene a ed, as pa ame ic
models in ol e a ma ginalisa ion o e combina ions o pa ame e s and hus implic-
i ly ely on unde lying p esump ions (Xie 2020). In con as , obse a ion-d i en
2762
M.Mu h e al.
“ML models do no make any p e-speci ied assump ions abou he unc ional o m
o he equa ion, he in e ac ion be ween a iables, and he s a is ical dis ibu ion o
pa ame e s” (Ghoddusi e al. 2019). In ac , ML algo i hms gene ally adop all da a
inpu o aining and a e hus able o i he model ac oss e y di e en da a s uc-
u es wi hou equi ing speci ic a p io i conside a ions (Xie 2020).
ML can be ca ego ised acco ding o i s associa ed lea ning s yles. Supe ised,
unsupe ised, and ein o cemen ML a e ecognised as gene al co e ca ego ies (E ns
e al. 2020; Mu phy 2012). In addi ion, u he lea ning s yles can be ound in he
esea ch li e a u e (see u he Ma and Sun 2020; Zhang 2020). In pa icula , supe -
ised ML is used o a a ie y o p ac ical p edic ion asks in a eas such as ma ke ing
(Ma and Sun 2020), o which eason i is desc ibed in mo e de ail. A majo goal o
supe ised lea ning is o d aw conclusions abou u u e o ye unknown de elop-
men s om exis ing in o ma ion. Fo his pu pose, one o mo e a ge a iables
y=(y1,…,y
n)
a e associa ed wi h a ce ain numbe o po en ial inpu a iables
x=(X1,…,X
n)
. Subsequen ly, a model is i ed in such a way ha he alues o he
inpu a iables co espond o he alues o he a ge a iable wi h a minimum e o :
∶x
→
y
. This p o ides he oppo uni y o in oduce new inpu a iables
x
=(
X
1
,…,
X
n)
in o he model and o de i e a p edic ion abou unknown a ge a i-
ables
y=(
y1,…,
yn)
o a bi a y alues
(
Xi
)
=y
i
. This can be applied o make a
p edic ion abou he classi ica ion o a g oup— o example, who will be a po en ial
buye —which is de ined as a disc e e p edic ion p oblem. In addi ion, a p edic ion o
me ic alues is also possible— o example, abou he expec ed u no e —which is
known as a me ic p edic ion p oblem (Good ellow e al. 2016; Mu phy 2012).
2.2 P edic i e modelling
Alb ech e al. (2021) unde line he ecen de elopmen o ML in he speci ic con-
ex o p edic ion, no ing ha while in he pas mo e adi ional s a is ical o u he
empi ical me hods we e employed, “mo e ecen ly, ML as a subse o AI has been
added o he domains con ibu ing e ec i ely o business p edic ion p oblems.” P e-
dic ion is one o he main applica ion p oblems o ML. Howe e , he e a e o he s,
such as he p ocessing o images o objec classi ica ion o na u al language p o-
cessing o speech gene a ion (Black e al. 2022)—including Gene a i e P e- ained
T ans o me s (GPT) such as Cha GPT (Esmaeilzadeh 2023), which a e beyond he
scope o his e iew pape . The e m “p edic ion” is equen ly used in business and
ma ke ing managemen in ega d o p ojec ions abou unknown u u e s a es based
on pas and ongoing da a (Kozak e al. 2021; Seyedan and Ma akhe i 2020). By
i s e y na u e, his e m he e o e implies a u u e-o ien ed pe spec i e on he ou -
come. In o de o dis inguish he ole o p edic ion om he closely ela ed e m
“ o ecas ”, Kmiecik and Zangana (2022) o e a di e en ia ion. They speci y ha
“ o ecas ing is a ype o p edic ion, and i bases he u u e ou comes on empo al
eco ded da a […] In essence, e e y o ecas is a ype o u u e p edic ion; howe e ,
no all u u e p edic ions a e o ecas s, as o ecas s ocus on no only a u u e occu -
ence bu also he ime o he occu ence.” Thus, when compa ing p edic ion and
o ecas ing, he au ho s emphasise o he la e he necessi y o empo ally eco ded
2763
The applica ion o machine lea ning o demand p edic ion…
da a, also known as a ime se ies, which is a sequence o ch onologically a anged
alues
(X1,…,X
n)
, measu ed o a ce ain ime
( 1,…,
n)
.
Taking a mo e de ailed iew o he ma e , a ious ypes o app oaches can be
conside ed, such as quan i a i e p edic ions which allow o a measu emen o he
p edic i e e o , and which a e he main ocus o his e iew. In addi ion, quali a i e
p edic ions can be applied, o ins ance, based on he in ui ion o expe s (Kmiecik
and Zangana 2022). Fu he mo e, he e a e di e en ypes in e ms o he numbe
o inpu a iables, also called p edic o s, whe eby ei he one (uni a ia e) o mul-
iple (mul i a ia e) p edic o s can be included in a model (Hombu g 2020). In his
espec , Seyedan and Ma akhe i (2020) emphasise he ele ance o da a inpu s ha
can p o ide addi ional explana o y po en ial o p edic ing u u e e en s. They sug-
ges : “Inco po a ing exis ing d i ing ac o s ou side he his o ical da a, such as eco-
nomic ins abili y […] could help adjus he p edic ions wi h espec o unseen u u e
scena ios o demand.” This sugges s ha a demand beha iou analysis equi es
conside a ion o he impac exe ed by he p e ailing ci cums ances. Fo example,
demand beha iou can be in luenced by company- ela ed mic o- ac o s, such as p o-
mo ional ac i i ies. In addi ion, ex e nal mac o- ac o s, such as ma ke condi ions,
shape he en i onmen in which demande s ope a e (A un aj and Ah ens 2015; E a
e al. 2022).
2.3 Demand unde ola ili y
The e m “ ola ili y” is employed in pa icula in econome ic analyses, as well as in
he inancial sec o , and is commonly exp essed as he s anda d de ia ion (σ). This
me ic is he squa e oo o he a iance, de ined by he ollowing o mula o a
gi en popula ion da ase o x1, x2, …, xn:
σ=�
1
N∑
N
i=1�
xi−μ
�2
(N = o al numbe
o obse a ions; μ = mean; xi = i h obse a ion) (Mondello 2022). In o de o de e -
mine s anda d de ia ion on a compa able basis o ola ili y measu emen , i is ypi-
cally benchma ked agains a e e ence alue, wi h conside a ions o addi ional com-
ponen s such as end (Ca iolle and Goujon 2013; Loayza e al. 2007; Raju and
Acha ya 2020). Howe e , i should be emphasised ha ola ili y can also be unp e-
dic able, especially when causal ela ionships o ci cums ances we e p e iously
unknown o no e iden (Angus e al. 2023). Demand is one o he a eas cha ac e -
ised by ola ili y, as Abolghasemi e al. (2020) no e: “ he demand o a pa icula
p oduc o se ice is ypically associa ed wi h di e en unce ain ies ha can make
hem ola ile and challenging o p edic ”, making i a c i ical aspec o ma ke ing
p edic ions. Po en ial s a egies ha migh be conside ed o cope wi h ola ili y in
demand include inc easing in en o y le els o capaci y o co e he luc ua ion.
Howe e , his has di ec business implica ions in e ms o in en o y managemen ,
he cos s associa ed wi h o e s ocking and impac s on capi al commi men and
liquidi y (Lin e al. 2022). As a esul , hese s a egies can in ol e signi ican addi-
ional e o along a ma ke ing supply chain (Kmiecik and Zangana 2022). On his
poin , Kmiecik and Zangana (2022) indica e ha “demand luc ua ions could imply
supply managemen p oblems and c ea e a endency o keep excessi e s ocks as a
bu e o p oduc ion sys ems. Using lexible and p ecise o ecas ing p ocedu es
2764
M.Mu h e al.
gi es possibili ies o gain good esul s e en in cap icious ma ke s.” This esul s in
he need o deepen he unde s anding o how o an icipa e ola ile demand pa e ns
and inco po a e hem o a ce ain ex en in o p edic i e me hods o guiding
in o med decision-making (Lin e al. 2022).
In his con ex , a mo e de ailed obse a ion o he mac oeconomic ci cum-
s ances can p o ide insigh s, especially i hey exe a signi ican impac on demand
(Hasheminejad e al. 2022). This equi es looking beyond he pe spec i e o a single
company as an indi idual economic uni and inco po a ing he gene al mac oeco-
nomic en i onmen ha encompasses he o e all si ua ion o he eal and mone a y
economies (Con ad 2020). F om a eal-economy pe spec i e, he e is a sys ema ic
shi in he way consume s alloca e hei budge depending on economic de elop-
men s (Kamaku a and Du 2011). Kamaku a and Du (2011) explain ha du ing
economic down u ns, “consume s gene ally educe hei consump ion budge [ o
non-essen ial commodi ies] ei he because hei income is lowe , o because hey
become mo e isk a e se, alloca ing mo e o hei income owa ds sa ings, which
o ces hem o sa is y essen ial needs i s ”. Rela ed o his, cus ome s’ esponses o
p ice changes can be a ec ed by economic g ow h a es (Go don e al. 2013). Fo
example, in many p oduc ca ego ies, p ice sensi i i y inc eases when he economy
weakens (coun e cyclical beha iou ). Acco ding o Go don e al. (2013), his seems
“consis en wi h he in ui ion ha consume s become mo e p ice sensi i e du ing
weake economic pe iods.” Howe e , his is no necessa ily always applicable. In
ega d o subs i u e goods, o ins ance, demand may e en inc ease when a mac o-
economic down u n occu s (p ocyclical beha iou ). Hence, ola ile mac oeconomic
ci cums ances c ea e a dynamic en i onmen o demande s’ beha iou al esponses.
2.4 Mac oeconomic en i onmen
F om an academic iewpoin , Loayza e al. (2007) see ola ili y induced by he mac-
oeconomic en i onmen as being ela ed o ac o s such as ex e nal shocks, eco-
nomic policies o mic oeconomic and ins i u ional dis o ions. In ega d o he cu -
en si ua ion, Taskan (2022) no es: “In ecen yea s, o ganisa ions ha e aced la ge
and unexpec ed e en s, such as inancial c ises, he COVID-19 pandemic, clima e
change and wa , wi h a la ge impac on he wo ld a se e al economic and socie al
le els, and he ac onym VUCA [Vola ili y, Unce ain y, Complexi y und Ambigu-
i y] has been equen ly used by schola s and p ac i ione s o y o unde s and such
en i onmen al dynamics.” The ecen example o COVID-19, wi h i s majo impli-
ca ions, is also aken up by Tudo (2022) and classi ied as a “black swan” e en ,
e e ing o i as an ex e nal shock o he gene al en i onmen wi h se e al mul i-
laye ed consequences. The associa ed e ec s on he beha iou o demande s a e
e lec ed, o example, in he inc eased con idence in e als o demand pa e ns du -
ing his pe iod (Ahmed e al. 2022; Ma and Fildes 2020).
Ano he e y ecen and o some ex en ela ed example o he shi ing cu en
mac oeconomic landscape is he eme gence o high in la ion a es, in e wined
wi h changes in mone a y policy, shaping he beha iou o demande s. The e m
2765
The applica ion o machine lea ning o demand p edic ion…
“in la ion” he e is con en ionally unde s ood as a sus ained inc ease in he gene al
p ice le el (Con ad 2020). Fo example, annual in la ion in he Eu ozone was a a
ema kable le el o 8.4% in 2022 and is p ojec ed by he Eu opean Cen al Bank
(2023) o emain abo e i s 2% a ge un il 2025. The eason o his high in la ion
is mul i-causal, and he de e mining ac o s include p ice inc eases in he ene gy
and ood sec o s. These a e in u n caused by ci cums ances such as he wa in
Uk aine—mos ecen ly, also he con lic s in he Middle Eas —bu also by ongoing
dis up ions in supply chains as a consequence o he COVID-19 pandemic (Minis-
y o Economic A ai s and Clima e 2023). Fo consume s, he occu ence o high
in la ion usually implies ha a gi en nominal amoun o money can be used o buy
ewe se ices and goods, which—acco ding o gene al economic unde s anding—
can lead o a loss o pu chasing powe ollowing highe p ices and he shi ing o
demand o goods and se ices (Sie e ing 2021). In addi ion, he cu en p ice may
send ambiguous signals abou u u e p ice de elopmen s in an in la iona y en i on-
men , which can lead o ola ile o misleading demand decisions. A u he compli-
ca ing ac o is ha he associa ed p ice inc eases usually do no occu a he same
ime bu in a delayed sequence (Con ad 2020). Hence, consume demand ypically
e lec s no only he p ice o an indi idual i em bu also i s p ice ela i e o he a e -
age o he ca ego y and, mo e b oadly, o he a e age o i s ca ego y ela i e o o he
ca ego ies (Danahe and B odie 2000). Howe e , no only do he ac ual p e ailing
condi ions ma e bu also he expec a ions o households abou he u u e de el-
opmen o he mone a y si ua ion, as well as hei own pe cei ed unce ain y. As
shown in he esea ch by Duca e al. (2010), consume s a e mo e likely o make a
la ge pu chase when a la ge change in in la ion is expec ed. As a conclusion o his
discussion, i appea s almos impe a i e o conside he mac oeconomic si ua ion
and i s e ec s in g ea e de ail when dealing wi h demand p edic ions.
3 Resea ch me hodology o heli e a u e e iew
As he academic con ibu ion o a li e a u e e iew is essen ially de e mined by he
compelling quali y o i s me hodological design, Linnenluecke e al. (2019) posi
ha he employmen o “ igo ous me hods and he cla i y o epo ing, as well as on
he applica ion o scien i ic s a egies” a e c ucial componen s o a sys ema ic li e a-
u e e iew. Simila ly, Wal e (2021) indica es ha only “a well-s uc u ed, anspa -
en , and eplicable me hodology esul s in a eliable basis o knowledge” (see u -
he Fink 2014; Fisch and Block 2018; Pe ic ew and Robe s 2006). Fo his eason,
his e iew pape con ains a desc ip ion o i s exac me hodology, which encom-
passes an o e iew o he s uc u al app oach and a sys ema ic explana ion o he
indi idual s eps. The e o e, his sys ema ic li e a u e e iew di e s in i s app oach
o analy ical li e a u e e alua ion om ial-and-e o me hodologies (Kaushal e al.
2023; T an ield e al. 2003).
Adhe ing o he amewo k p o ided by Xiao and Wa son (2019) o a sys em-
a ic li e a u e e iew, his me hodology gene ally con ains h ee main phases,
which a e de ined as: planning he e iew, conduc ing he e iew, and epo ing he
2766
M.Mu h e al.
e iew (see Fig.1). These s ages show a high deg ee o simila i y wi h hose o
T an ield e al. (2003), which a e widely used in he cu en managemen li e a u e
(Du ugbo and Al-Balushi 2022; El Shoubaki e al. 2021; Göcke e al. 2021). None-
heless, Xiao and Wa son’s (2019) sugges ion is ollowed in his e iew pape , as i
inco po a es mo e ecen scien i ic indings and p o ides a comp ehensi e no a ion
and de ailed desc ip ion o each subca ego y. The i s main phase o planning he
e iew includes wo sub-s eps. The i s sub-s ep in ol es o mula ing he o e a ch-
ing esea ch p oblem o “How can ML be e ec i ely applied o p edic demand pa -
e ns unde mac oeconomic ola ili y?” which, as p e iously explained, aligns wi h
he s uc u e o he “CIMO” logic. The second sub-s ep equi es he de elopmen
and alida ion o a e iew p o ocol, which ep esen s a p ede ined plan de ailing he
esea ch me hodology. As depic ed in Fig.1, he e iew p o ocol se es o ensu e
he eliabili y o he en i e s udy by enabling o he s o eplica e and e i y he esul s
Fig. 1 Re iew p o ocol (adap ed om Xiao and Wa son 2019)
2773
The applica ion o machine lea ning o demand p edic ion…
signi ican con ibu ion o he chu n p edic ion and he deep lea ning keywo ds. In
con as , “Ene gy” (n = 53) and “Applied Ene gy” (n = 49) a e pa icula ly high-
con ibu ing jou nals o he keywo d o ecas ing. Fu he mo e, on he igh -hand
side o Fig.4, i can be seen ha he a icle by B eiman (2001) is men ioned mos
equen ly in connec ion wi h he keywo d ML and is ci ed a o al o 16 imes as
e e ence. In addi ion, he con ibu ions o F iedman (2001) (n = 7), Coussemen and
Poel (2008) (n = 6) and Neslin e al. (2006) (n = 5) a e ci ed wi h high equency. Fo
he keywo d o ecas ing, Hyndman and A hanasopoulos (2018) (n = 6) and Hynd-
man and Koehle (2006) (n = 5) eme ge as e y impo an . Finally, i is e iden om
he sankey diag am ha he leading sou ce o he keywo d demand o ecas ing is
he “In e na ional Jou nal o P oduc ion Economics” (n = 53), wi h a p onounced
emphasis on echnical pe spec i es, while o he keywo d sales o ecas ing, he p i-
ma y sou ce is he jou nal “Managemen Science” (n = 48), wi h a s onge busi-
ness-o ien ed ocus. Consequen ly, u he a en ion is needed o moni o how ce -
ain e minologies o simila opics a e mo e p e alen in speci ic academic (sub)
disciplines han in o he s, pa icula ly in ega d o in e disciplina y s udies.
4.2 Analysis o abs ac s
In his sec ion, an examina ion o abs ac s o all publica ions in he li e a u e
sample (n = 64) is p o ided o he pu pose o syn hesising co e in o ma ion and
Flowcha : jou nals -keywo ds-sou ces
Rega ding he li e a u e sample
jou nals keywo ds e e ences
Fig. 4 Sankey diag am o jou nals—keywo ds— e e ences. No e: Figu e 4 isualises a “Th ee-Fields
Plo ”, wi h “Le Field” = Jou nals (Sou ces), “Middle Field” = Keywo ds (Au ho s’ Keywo d) and
“Righ Field” = Re e ences (“Ci ed Re e ences in he pape s”), using he pa ame e se ing “Numbe o
I ems = 7” (A ia and Cuccu ullo 2023; A ia and Cuccu ullo 2017). The sankey diag am is limi ed o he
mos common jou nals, keywo ds, and e e ences, and can hus o e simpli y he complex in e connec-
ions be ween hem. Only he i s -named au ho o he e e ence is men ioned on he igh -hand side o
acili a e a concise isualisa ion
2774
M.Mu h e al.
disco e ing s uc u al ela ionships among hem (T app 2012). Fo his, a concep-
ual s uc u e map (see Fig.5) is gene a ed o iden i y he speci ic knowledge p o-
ile in he abs ac s (A ia and Cuccu ullo 2017). A he ou se , ele an e ms a e
il e ed om all abs ac s, emo ing i ele an e ms such as de e mine s o con-
junc ions om he o al se o wo ds and applying Po e ’s (1980) s emming algo-
i hm o educe all wo ds o hei oo o m. Then, a Co espondence Analysis is
applied as a da a educ ion echnique o map he hema ic spaces in wo dimen-
sions, ollowed by K-Means Clus e ing o iden i y clus e s o s udies exp essing
common concep s (A ia and Cuccu ullo 2017). Wi hin he abs ac s, i e ocal
poin s (F1-5) can be iden i ied, which a e assigned a supe o dina e desc ip ion
e m. The la ges one (blue) can be summa ised unde he ca ego y (F1) ML
me hods and hei applica ion con ex , which is, in addi ion o me hodological
e minologies (e.g., algo i hms, models), ela ed o he ocused a ea. This seems
app op ia e, conside ing ha o e 350 algo i hm o model usages a e egis e ed
in he s udies o he li e a u e sample. I is also appa en in he concep ual map
ha ano he ocal poin (b own) exis s ha deals wi h he (F2) Economic ac o s.
This can be a ibu ed o he ac ha a numbe o s udies in ol e conside a ion
o economic impac , ma ke ci cums ances, sen imen s, o mac o-en i onmen
Fig. 5 Concep ual s uc u e map. No e:Figu e5 isualises a “Wo d Map” ia he “Concep ual S uc-
u e” command based on a wo d occu ence ma ix and wi h he pa ame e se ing “Me hod = Co e-
spondence Analysis” and “Field = Abs ac s”. In heK-Means Clus e ing, allclus e s wi h ewe han 3
wo ds we e emo ed o cla i y pu poses (A ia and Cuccu ullo 2023; A ia and Cuccu ullo 2017)
2775
The applica ion o machine lea ning o demand p edic ion…
in hei da ase s o in hei p edic i e modelling app oaches. Two u he ocal
poin s co e ed in he abs ac s a e o a mo e echnical na u e. They in ol e (F3)
Time se ies o ecas ing ( ed) and he usage o (F4) Neu al ne wo k p edic ions
(g een). While 27 s udies ocus ei he en i ely o pa ially on ime se ies da a o
hei empi ical modelling, A i icial Neu al Ne wo ksa e employed 89 imes o
his pu pose—some imes in a ious ways wi hin a single s udy—which unde -
lines he ele ance o his me hod. The emaining ocal poin ela es o (F5) Cus-
ome amewo ks (o ange), as cus ome - ela ed ac i i ies and p osseses a e a
consis en esea ch mo i e o ML applica ions in his speci ic esea ch con ex .
4.3 Analysis o ull ex s
The analysis o he abs ac s is ollowed by a e iew o he ull ex s, whe e he
aim is o p o ide a opic-cen ed o e iew and ecapi ula e he key ex ual pe spec-
i es. The app oach adhe es o some o he ins uc ions p o ided by Linnenluecke
e al. (2019), who ecommends esea che s going “ h ough ways ha p io publi-
ca ions ha e con ibu ed o de eloping […] unde s anding o hemes, concep s o
phenomena o in e es .” This p ocess is an a emp o ind possible app oaches on
how o managing he ML applica ion p ocess, adap ing i o mac oeconomic ola ile
condi ions and he in e disciplina y equi emen s o p edic ing demand beha iou .
To ensu e concise epo ing o he indings, hey a e p o ided in acco dance wi h
he common ch onological o de along he ML wo k lows, using he ollowing h ee
o e a ching ca ego ies: p ep ocessing, modelling, and pos p ocessing. This includes
unde lying ques ions such as which s a e-o - he-a model algo i hms a e mos
sui able in his con ex and which (mac oeconomic) p edic o s a e app op ia e o
accu a e demand p edic ions. Beyond summa ising he majo indings o he pape s
wi hin hose subca ego ies, he e iew also in ol ed an a emp o iden i y di e gen
iews and inconsis en di ec ions o esea ch app oaches (T app 2012).
4.3.1 P ep ocessing
The ollowing sec ion con ains an analysis o he p ep ocessing s a egies ha
aim o enhance ML esul s in p edic ing demand in he ma ke ing con ex wi hin
dynamic mac oeconomic se ings. P ep ocessing is p edominan ly ocused on
collec ing aw da a as well as he subsequen handling p ocedu e o his da a o
cons uc an op imal ea u e se o he p edic i e model (Kha an e al. 2021;
Punia and Shanka 2022; Wang 2022). Au ho s desc ibe his s age as bo h a chal-
lenge and a c i ical de e minan o he inal quali y o he model, wi h a p ima y
ocus on selec ing an app op ia e se o inpu a iables (Raizada and Saini 2021).
This sec ion includes a compa ison o he pe o mances o uni a ia e models ha
ely on a single inpu a iable wi h complex models ha ac o mul iple a iables,
wi h an emphasis on hei espec i e p edic i e capabili ies in ola ile mac o-
en i onmen se ings. Pa icula a en ion is gi en o he selec ion o model a i-
ables o an icipa ing he mac oeconomic condi ions.
2776
M.Mu h e al.
In ac , a la ge numbe o he e iewed s udies in he li e a u e sample (n = 49)
ei he pa ially o en i ely used empi ical models which include mul iple inpu
a iables (mul i a ia e). Wi hin he con ex o demand p edic ion, nume ous
pape s highligh he ad an age o mul i a ia e me hods compa ed o uni a i-
a e app oaches (e.g., Abolghasemi e al. 2020; Cla e ia e al. 2020; Punia and
Shanka 2022). This p e e ence p ima ily seems o be a ibu able o he inhe -
en cons ain s o uni a ia e me hods ha ely solely on a single inpu a iable,
such as his o ic demand beha iou , o ex apola e u u e de elopmen s. Punia
and Shanka (2022) measu e in hei empi ical in es iga ions supe io pe o -
mance by employing he same p edic ion model wi h a he han wi hou con-
ex ual a iables. Consequen ly, Abolghasemi e al. (2020) conclude ha speci ic
uni a ia e me hods “wo k only well when he u u e is simila o he pas […]
[and] migh ail o o ecas well i demand ime se ies is subjec o ola ili y.”
Simila ly, Hasheminejad e al. (2022) unde sco e ha adi ional, a he uni a i-
a e p edic ion me hods “gene ally do no wo k when he ma ke is cons an ly luc-
ua ing.” This pe spec i e aligns wi h he unde s anding ha demand is ypically
in luenced by a mul i ude o ac o s (Ma and Fildes 2020), he eby implying ha
he in eg a ion o ex e nal ac o s holds he po en ial o make a numbe o unce -
ain ies in u u e demand p edic ions explainable, hus educing hem (Ghod-
dusi e al. 2019). The empi ical s udies in he li e a u e e iew u he suppo
he impo ance o a comp ehensi e in eg a ion o inpu a iables. An analysis o
he s udies in his e iew e eals ha esea che s conside a median o 14 inpu
a iables o hei modelling, wi h he maximum obse ed in he la ge cus ome
da ase used by Wang e al. (2019) wi h a o al o 898 inpu a iables (see Fig.6).
Rega ding he speci ic con ex o his esea ch, Wang (2022) di e en ia es
he p oblem o selec ing app op ia e model a iables in o wo dis inc issues: (1)
de e mining whe he and how he a iables exhibi empo al causali y ( ime lags)
Basedon he da ase s o heempi ical s udiesin he li e a u e sample
Mean
56.5
Median
14.0
SD
142.5
Min
1.0
Max
898.0
25 h Pe cen ile
6.0
75 h Pe cen ile
31.5
Fig. 6 Numbe o inpu a iables. No e:I a s udy con ains mul iple da ase s, each one is accoun ed o
in he e alua ion, which may esul in a single s udy con ibu ing mo e han one se o inpu a iables.
Figu e6 is based on he da ase s o he empi ical s udies
2777
The applica ion o machine lea ning o demand p edic ion…
and (2) iden i ying he mos e ec i e p edic o s. Conce ning he i s issue (1),
he au ho classi ies po en ial inpu a iables based on hei ime dependencies
in o h ee ca ego ies: leading, coinciden , and lagging a iables. These ca ego ies
indica e whe he a po en ial inpu is expec ed o change be o e, a e , o simul a-
neously wi h a a ia ion in he demand ou pu . When p edic ing in he con ex o
b oade economic condi ions, Wang (2022) highligh s he signi icance o he i s
one and posi s ha a “leading indica o helps […] p edic u u e changes be o e
he ou come o he economy begins o go up o down [and can hus] used as
an ale signal.” Poza and Monge (2023) concu wi h his iew, emphasising he
ad an age o leading a iables in no only an icipa ing ends bu also disce ning
u ning poin s in he economic ma ke en i onmen .
Fo he p ac ical iden i ica ion o ele an demand dependencies, he e iewed
pape s con ain men ions o he Au oco ela ion Func ion o he Pa ial Au oco -
ela ion Func ion (n = 5) as a ele an app oach o iden i y hem be o e model-
ling. These me hods indica e he co ela ion o a po en ial demand se ies wi h i s
delayed a ian s, espec i ely hei esiduals; hence, hey p o ide guidance o
de e mining he lag s uc u e wi hin he da a (see u he Alsah e e al. 2022;
Bukha i e al. 2020; Con e as-Masse e al. 2022; Kmiecik and Zangana 2022;
Meisenbache e al. 2022). O he me hods a e also employed by esea che s, such
as Wang (2022) and Wu e al. (2022), using Akaike and Schwa z In o ma ion C i-
e ia, while some au ho s do no explici ly indica e hei p ocedu e o de e min-
ing he lag s uc u e o conside pe o ming his immedia ely wi hin modelling
(e.g., Gü ses-T an and Mon i 2022; Liu e al. 2021).
To ga he da a encompassing such leading a iables o demand p edic ion,
se e al s udies sugges ele an app oaches (e.g., Ghonghadze and Lux 2012;
Pe opoulos and Siakoulis 2021), such as using u u e-o ien ed sen imen s o
expec a ions o an icipa e he mac oeconomic si ua ion. Since hey a e no
di ec ly obse able, wo majo app oaches can be iden i ied in he li e a u e sam-
ple o ob ain u he in o ma ion abou his: a mo e adi ional app oach ha uses
economic indica o s, especially based on sen imen da a om su eys, and a ela-
i ely mode n app oach elying on sea ch engine ends. Economic indica o s can
be de i ed o ins ance om economic endency su eys ha add ess ques ions
such as how indi iduals pe cei e hei own inancial si ua ion o he economy in
gene al. Using su ey esponses is seen o be a alid sou ce o in o ma ion o
his pu pose, as: “(a) hey a e based on he knowledge o agen s ha de ac o
ope a e in he ma ke , (b) hey con ain in o ma ion on a wide a ie y o economic
a iables, and (c) hey a e a ailable p io o he publica ion o o icial da a”
(Cla e ia e al. 2020). Responden s a e asked o e alua e whe he such economic-
ela ed de elopmen s a e likely o de elop posi i ely, emain unchanged, o
de elop nega i ely in he u u e. F om he esponses o a su ey, balances can be
compu ed by con as ing posi i e and nega i e esponses, excluding neu al ones,
and di iding by he o al numbe o esponden s
(
Balance
(B)=
Posi i e esponse (n+)− Nega i e esponse (n−)
numbe o pa icipan s
(
n
o al)
) (Ghonghadze and Lux 2012;
Cla e ia e al. 2020). The economic sen imen indica o s de i ed a e conside ed
as “key o moni o ing he cu en s a e o he economy and p o iding o wa d
2778
M.Mu h e al.
looking in o ma ion” (Cla e ia e al. 2020), and Meisenbache e al. (2022)
obse e ha sales da a in pa icula exhibi a equen co ela ion wi h such eco-
nomic indica o s.
In e ms o he speci ic empi ical applica ion o pa icula indica o s, no uni e -
sally adop ed indica o s could be iden i ied in he li e a u e sample. This sugges s
ha he indica o s o economic sen imen can also be ambiguous, hei usage ends
o be complex, and none o hem is consis en ly applicable ac oss all demand p edic-
ion scena ios in ma ke ing, making i impo an o c i ically e alua e hese indica-
o s on a case-by-case basis. Howe e , he e a e ce ain s udies ha e e o o icial
indica o s associa ed wi h he Eu opean Union, o example in he con ex o Join
Ha monised EU P og amme o Business and Consume Su eys. Cla e ia e al.
(2020) discuss he co esponding consume con idence and indus y con idence
indica o s, whe e he o me e lec s expec a ions ega ding employmen , expo , o
p oduc ion and he la e e e s o gene al economic condi ions o speci ic demand
decisions in he nex 12mon hs. In addi ion, Ghonghadze and Lux (2012) deal wi h
economic sen imen indica o s based on su ey da a om he Eu opean Commis-
sion, ocusing on speci ic sec o s such as cons uc ion, consume , manu ac u ing,
e ail and se ices. Fu he mo e, he pape by Pe opoulos and Siakoulis (2021) is
pa icula ly ele an o cap u ing ola ili y o his esea ch backg ound and deals
wi h he ola ili y index VIX. This index is based on he luc ua ions o he S&P
500 Index, ep esen ing he 500 la ges lis ed companies in he U.S., and e lec s i s
expec ed luc ua ion in he coming mon h (see u he Leh e e al. 2021). In hei
s udy, Pe opoulos and Siakoulis (2021) also apply his S&P 500 index o ope a-
ionalise a “c isis e en ”, which hey desc ibe as a scena io ha sees i declining by
mo e han 8% o e a 3-mon h pe iod. This de ini ion p o ides an objec i e measu e
o signi ican ma ke ola ili y and can hus con ibu e o i s inco po a ion in o quan-
i a i e ML p edic ions.
In ela ion o collec ion o ele an de e minan s o demand, many au ho s in he
li e a u e sample also ecommend al e na i e indica o s o e su ey da a, pa icu-
la ly emphasising ad anced sea ch engine ends (e.g., Punia and Shanka 2022;
Tsao e al. 2022; Tudo 2022). Poza and Monge (2023) sugges he use o sea ch
que ies, which can show a co ela ion wi h key economic indica o s o p o ide p e-
dic i e insigh s abou u u e di ec ions and Tudo (2022) u he subs an ia es his,
summa ising wi hin he p esen esea ch con ex ha “ he inclusion o GT [Google
T end] in o ma ion o e s signi ican bene i s in he o m o imp o ed o ecas ing
pe o mance. None heless, GT [Google T end] da a has also been acknowledged in
p e ious esea ch as a leading indica o o key a iables o in e es .” In ligh o
exis ing esea ch, Ghoddusi e al. (2019) highligh he po en ial o in e ne sen imen
o ola ili y- ela ed p edic ions and conside ML echniques o be a powe ul ool
o his pu pose. In addi ion o o he s udies, Ryu e al. (2020) obse e: “P edic ion
o economic ac i i ies by using social ne wo k da a o in e ne sea ch da a ahead
ac ual ac i i ies has been epo ed in he s ock ma ke , ma ke ing and ou ism.”
Fo his, he ocus is ypically on he popula i y o speci ic keywo ds, e lec ed in a
sea ch engine end index ha indica es he amoun o que ies in ce ain geog aphi-
cal a eas ei he in absolu e numbe s o no malised e ms (Punia and Shanka 2022;
Tudo 2022). As Punia and Shanka (2022) no e, sea ch que ies o igina ing no only
2779
The applica ion o machine lea ning o demand p edic ion…
om Google bu also om o he pla o ms such as YouTube, and e lec ed in a spe-
ci ic index, can also indica e a posi i e ela ionship in demand p edic ions.
Fu he mo e, an impo an obse a ion o men ion is ha he li e a u e also co -
e s a b oad spec um o o he p edic i e a iables o demand. They encompass a -
ious dimensions such as consume beha iou , demog aphics, wea he , ad e isemen
measu es, p oduc da a, o seasonal and calenda e ec s.2 The p ice a iable is in
his con ex a u he c i ical ac o in academic discou se a ound he in es iga ed
esea ch ques ion: “The impac o p ice changes should no be igno ed while design-
ing algo i hms o p edic ing cus ome choice” (Chen e al. 2021). Beyond i s inhe -
en ele ance, p ice can also indi ec ly e lec mac oeconomic condi ions like in la-
iona y en i onmen s, po en ially possessing p edic i e alue o b oade economic
ci cums ances. Menhaj and Ka oosi-Kalashami (2022) emphasise he complexi y o
he p ice a iable in his con ex , no ing ha p ice changes can occu due o mac o-
economic condi ions bu also due o seasonal, cyclical, o end componen s. Such
p ice luc ua ions may, in u n, lead o a ying p edic i e implica ions, u he high-
ligh ing he need o sophis ica ed unde s anding and applica ion o he p ice a i-
able in demand p edic ions du ing mac oeconomic ola ili y.
A e conside ing empo al causali y and especially leading indica o s, issue (2)
desc ibed by Wang (2022) is ele an , which in ol es iden i ying he mos e ec i e
inal se o po en ial p edic o s o he model. The eason o ocusing on he mos
impo an a iables is d i en, among o he easons, by he need o educe compu-
a ional ime and complexi y o he p edic i e model (Cas illo e al. 2017; Quin e o
e al. 2022). Such la ge da ase s a e also e lec ed in he li e a u e sample, whe e
no able examples include Ma ínez-de-Albéniz e al. (2020), who p ocess 1.74
illion pieces o clicks eam da a om an online lash sales e aile , and Bi e al.
(2022), who ackle 165 million sales ansac ions. Fu he mo e, ega ding he as
numbe o po en ial a iables, he s udy by Wu and Li (2021) ini ially conside 625
po en ial inpu a iables om he inancial sec o o p edic ing cus ome chu n, and
Hasheminejad (2022) s a s wi h 313 inpu a iables.
To add ess his issue, a ious me hods a e applied o ind an op imal subse o
ea u es ha p o ides he bes p edic i e esul s o he model. This ask is o en pe -
o med be o e he ac ual modelling bu is some imes also in eg a ed in o he model-
ling p ocess i sel (Meisenbache e al. 2022; Quin e o e al. 2022). Cas illo e al.
(2017) di e en ia e be ween wo p ima y me hods o p ocessing ea u es: ea u e
educ ion and ex ac ion. The aim o ea u e educ ion is o educe he o al numbe
o inpu a iables o he mos c i ical ones, while ea u e ex ac ion ans o ms he
exis ing ea u es and po en ially c ea es new o adap ed ones (Cas illo e al. 2017).
2 Examples o consume beha iou include click a e o paymen his o y and o demog aphics gende
o educa ion le el (e.g., Balles a e al. 2019; Chen 2022; Esmeli e al. 2021; Pu e man e al. 2020). Fo
wea he , examples a e empe a u e o wind speed (e.g., Liu e al. 2021; Punia and Shanka 2022). Ad e -
isemen da a encompasses campaigns o discoun s (e.g., Punia and Shanka 2022; Wang e al. 2019).
Examples o p oduc da a a e a icle ca ego ies and a ibu es (e.g., Kha an e al. 2021). Seasonal and
calenda e ec s e e o aspec s such as holidays and weekdays (e.g., Alb ech e al. 2021; Ma and Fildes
2020; Raizada and Saini 2021).
2780
M.Mu h e al.
Fo educing he se o a iables o he mos ele an demand ea u es, a numbe
o supe ised lea ning p ocedu es om adi ional s a is ical analysis can be iden-
i ied in he li e a u e sample as being ele an o his pu pose. These consis o
eg ession me hods (n = 7), including no only simple linea eg ession bu also o he
a ian s o eg ession, such as he Leas Absolu e Sh inkage and Selec ion Ope a o
(LASSO) me hod (e.g., Meisenbache e al. 2022; Pu e man e al. 2020). Co ela ion
analysis also plays an impo an ole (n = 7), o example measu ed by Pea son’s co -
ela ion coe icien (e.g., Gü ses-T an and Mon i 2022; Liu e al. 2021; Shaikh e al.
2022). In addi ion o pu ely quan i a i e app oaches, he e a e also s udies employ-
ing expe knowledge, like ha ob ained ia su eys o judgmen s om specialis s
in ele an academic o p o essional ields o by esea che s hemsel es (n = 6), wi h
he aim o selec ing impo an ea u es (Cas illo e al. 2017; Gü ses-T an and Mon i
2022; Hasheminejad e al. 2022; Khandani e al. 2010; Miloše ić e al. 2017; Tsao
e al. 2022).
Fo ea u e ex ac ion, in con as , unsupe ised me hods ha e a mo e impo an
ole, including p incipal componen analysis, examined in mul iple s udies (n = 7)
wi hin he e iewed li e a u e (e.g., Quin e o e al. 2022; Wu and Li 2021), and ac-
o analysis, used in wo cases (Poza and Monge 2023; Shaikh e al. 2022). Fu -
he mo e, he e a e also p ocedu es ha go beyond pu ely quan i a i e analysis. Fo
example, he inpu da a can also be ans o med o align wi h es ablished heo e ical
cons uc s o concep ual amewo ks, o ins ance om disciplines like economics
o ma ke ing. This can also be obse ed in s udies (n = 5) in he li e a u e sample,
such as by ans o ming a iables in o concep e sions o he Recency, F equency
and Mone a y alue (RFM) ega ding consume s’ demand o model hei beha iou
(e.g., Chashmi e al. 2021; Wu and Li 2021; Xie 2020).
4.3.2 Modelling
This sec ion examines he modelling phase, ocusing on selec ing and gene a ing
he mos e icien models o p edic ing in a ola ile mac oeconomic en i onmen .
The e o e, he sec ion begins wi h an in es iga ion o he gene al ele ance and sui -
abili y o ML algo i hms wi hin his speci ic esea ch con ex . This is ollowed by
an analysis o indi idual models, e alua ing hei usage o e ime and iden i ying
he models ha show he bes esul s in he li e a u e sample. Addi ionally, di e en
app oaches o e ec i e model selec ion and aining a e discussed o accoun o
mac oeconomic ola ili y.
The modelling sec ion is ini ially in ended o in es iga e he gene al ele ance
and sui abili y o ML o coping wi h ola ili y in demand p edic ion. In he majo -
i y o he s udies analysed in he li e a u e sample, ML models a e ound o pe o m
supe io ly and a e he e o e gene ally conside ed an impo an p edic ion app oach
in complex mac oeconomic con ex s. Fo example, he esea ch by Ma and Fildes
(2020) demons a es ha ML modelling (G adien Boos ing T ees) ou pe o ms
mo e adi ional me hods (e.g., Exponen ial Smoo hing, ARIMA, Naï e, The a) as i
is able o imp o e he an icipa ion o non-linea pa e ns in di e si ied da ase s. The
s udy by Alb ech e al. (2021) also illus a es ha ML modelling (Random Fo es )
has a supe io pe o mance compa ed o adi ional o ecas ing me hods (e.g., STL
2781
The applica ion o machine lea ning o demand p edic ion…
decomposi ion wi h ARIMA o ETS), which he au ho s la gely a ibu e o i s abil-
i y o an icipa e complex in e ac ions in ma ke ing- ela ed da ase s. Ano he exam-
ple o he applica ion o ML in dynamic ma ke condi ions is p o ided by he s udy
o Punia and Shanka (2022), who p esen a combined model o wo ML algo i hms
(LSTM, Random Fo es ). I shows supe io p edic i e pe o mance o sho -,
medium-, and long- e m demand p edic ion compa ed o adi ional app oaches
(e.g., ARIMA, eg ession) and mos e ec i ely inco po a es a la ge numbe o a i-
ables— om p omo ions o social media and wea he condi ions o egional eco-
nomic ac o s.
Howe e , in some o he s udies analysed, ypical ML me hods also p o e no
o be supe io . Abolghasemi e al. (2020) no e: “simple s a is ical models can ou -
pe o m some o he sophis ica ed ML and s a is ical models.” In he s udy by Lo i
e al. (2023), he de eloped ac ional calculus model—a gene alized di usion
model called GDMR—pe o ms mo e accu a ely han ypical ML models in cap u -
ing he dynamics o epea ed pu chases in p oduc li ecycles. In he con ex o he
ola ili y induced by he COVID-19 pandemic, Tudo (2022) examines he demand
o ideo con e encing solu ions and inds ha he adi ional s a is ical me hod o
exponen ial smoo hing pe o ms bes compa ed o classical ML algo i hms (e.g.,
A i icialNeu al Ne wo ks). S udies also show ha ML alone can be in e io o a
combina ion wi h o he me hods. This is illus a ed by he example o Pu ohi e al.
(2021), whose ag icul u al p ice p edic ions yield he bes esul s using ML (e.g.,
Suppo Vec o Machines) in mul iplica i e combina ion wi h s a is ical ime se ies
models (e.g., Exponen ial Smoo hing S a e Space Model).
Consequen ly, when aced wi h ola ile ma ke en i onmen s and complex and
non-linea demand dynamics, con en ional p edic ion echniques can each hei lim-
i s and ML solu ions p o ide a aluable con ibu ion. Al hough he e a e a signi ican
numbe o s udies in he li e a u e sample ha a ibu e supe io pe o mance o ML,
he e is no gene al p edominance. O he s—whe he adi ional s a is ical, ime se ies,
econome ic, o al e na i e ma hema ical-quan i a i e models—can also be supe io
in his con ex , necessi a ing indi idual es ing o he modelling algo i hms.
When i comes o selec ing a sui able se o algo i hms o he p edic ion p ob-
lem, he e is a wide ange o a ailable ML lea ning models o conside . Rega ding
he o e a ching lea ning s yle used in his con ex , a la ge majo i y o he s udies
(n = 54) use ei he ully o pa ially supe ised ML o empi ical modelling. In con-
as , o he lea ning s yles, such as unsupe ised o semi-supe ised lea ning, a e
each less commonly used (≤ 5) wi hin he empi ical co e models. Fo a de ailed
o e iew o he indi idual models used, in Fig.7 he numbe o model usages is
shown ac oss he empi ical s udies o e ime. Gi en he limi ed numbe o publica-
ions be o e 2019 and a e 2022 in he li e a u e sample, wi h less han h ee pe
yea , his analysis is ocused on his pe iod, whe e impo an de elopmen s a e e i-
den . Fu he mo e, beyond iden i ying he models commonly employed in s udies,
i is essen ial o highligh hose ha exhibi he bes o e all esul s and hence a e
ecommended o con inued applica ion in his con ex . To acili a e his, in Fig.8,
a g aphical ep esen a ion is p esen ed o he equency a which speci ic main ML
algo i hms a e ecognised by esea che s in he empi ical s udies o hei supe io
pe o mance. In Fig.8, a dis inc ion is highligh ed in e ms o he ype o p edic ion
2782
M.Mu h e al.
ou pu , since he selec ion o ML algo i hm can depend on his ac o , indica ing
o e all ha me ic p edic ion esul s a e gene ally conside ed mo e equen ly han
disc e e ones.
Addi ionally, i is impo an o no e ha Figs.7 and 8 p esen di e en models
dis inc ly, e en i hey a e pa o a combined app oach whe e mul iple algo i hms
a e used syne gis ically o op imise he o e all p edic i e ou come (Ghoddusi e al.
2019; Kha an e al. 2021; Wang e al. 2019). A p ima y objec i e behind his
app oach is o mi iga e he in luence o occasional subop imal demand p edic ions.
Fig. 7 Timeline o model usages
Fig. 8 P e e ed model by p edic ion ou pu ype. No e on Figs.7 and 8: A s udy can bo h use and ec-
ommend mul iple models—each model is hen coun ed indi idually. S udies ha do no ha e he con-
cep ual aim o using o e alua ing models, such as e iews, a e no conside ed in he igu es. A model is
assigned o he “o he ” g oup i i s usage coun is ewe han 10; i he usage coun is highe , he model
name is explici ly indica ed. Figu e7 is based on s udies om 2019–2022. Figu e8 has no ime pe iod
limi
2789
The applica ion o machine lea ning o demand p edic ion…
example, commonly used me ics may no adequa ely accoun o such domain-spe-
ci ic implica ions o he p edic ion e o .
Fo asks wi h disc e e p edic ion ou pu , he ROC (Recei e Ope a ing Cha ac-
e is ic) cu e and he A ea Unde his Cu e (AUC) ha e an impo an ole in he
li e a u e sample. The ROC cu e (n = 6) plo s he a e o ue posi i es agains he
a e o alse posi i es and hen he AUC (n = 10) shows he gene al pe o mance o
he ML model wi h ega d o dis inguishing be ween posi i e and nega i e classi-
ica ions (Chen e al. 2021; Geile e al. 2022). Rega ding he ou h and i h mos
equen me ics o his ou pu ype, which a e p ecision ( ue posi i e p edic ions
among all posi i e p edic ions) and ecall ( ue posi i e p edic ions among all ac ual
posi i e p edic ions), he esea che s Kozak e al. (2021) discuss u he implica ions
o hei use. Re lec ing on he balance be ween hese wo e o me ics in esponse
o changing economic condi ions, Kozak e al. (2021) explain: “Conside ing he
ola ili y o global alue chains, […], he need o lexible ma ke ing planning is
3.1
3.0
2.1
1.0
11.0
2.0
4.0
Fig. 10 Numbe o e o me ics. No e: E alua ion o he da ase s in he empi ical s udies
Basedon he da ase s o heempi ical s udiesin he li e a u e sample and il e edbyou pu ype
6
9
10
0510 15
ROC
Accu acy
AUC
Coun
E o me ic
Disc e e p edic ion ou pu
11
16
20
0510 15 20
MAPE
MAE
RMSE
Coun
E o me ic
Me ic p edic ion ou pu
Fig. 11 Mos equen ly used e o me ics. No e: E alua edby ou pu ype and based on he empi ical
s udies
2790
M.Mu h e al.
s onge han e e […] [ equi ing] o lexibly eac o as -changing condi ions, e.g.,
p io i izing ecall in chu n a ge ing when he economic clima e imp o es and p io -
i izing p ecision in di icul imes.”
When in e p e ing hese e o me ics, i is sugges ed in he li e a u e sample
o cons an ly ake in o accoun he demand p edic ion ho izon and i s associa ed
unce ain ies, especially when dealing wi h ola ile condi ions. Fo example, om
esea ch conduc ed by So jamaa e al. (2007), i is implied ha he e is a need o
e alua e p edic ion pe o mance in ela ion o he p edic ion ho izon. As ci ed by
Ma and Fildes (2020), hei wo k illus a es ha demand p edic ions in ended o p o-
ide an ou look o mul iple u u e pe iods (mul ile el p edic ions) pose a highe
le el o complexi y han single-pe iod demand p edic ions due o he accumula ion
o e o s, which in u n inc ease unce ain y. This ela ionship be ween he p edic-
ion ho izon, p edic i e pe o mance and unce ain y is documen ed by a ious
au ho s, including Kmiecik and Zangana (2022) o Wang e al. (2019). Addi ionally,
Alb ech e al. (2021) no e ha “o e all, he models’ pe o mances wo sen sligh ly
wi h inc easing lead ime.” Gi en his complexi y, Abolghasemi e al. (2020) a gue
o he de ini ion o p edic ion in e als, which speci y an uppe and lowe bound
o he p edic ions and ep esen he ange in o which p ojec ed demand is likely o
all a a ce ain con idence le el. They make he a gumen , ha such in e als can
en ich managemen unde s anding, s a ing: “P edic ion in e als will p o ide man-
age s wi h insigh s in o he mos app op ia e choice o o ecas ing me hods when
he deg ee o unce ain y is aken in o accoun ” (Abolghasemi e al. 2020).
Rega ding he second app oach, he b oade business-o ien ed e alua ion, he e
is an emphasis on he impo ance o looking beyond he me e measu emen o he
pe o mance o demand p edic ions. The con as is highligh ed by Gü ses-T an and
Mon i (2022), who dis inguish be ween “ o ecas e alua ion based on es ablished
o ecas e o s a is ics on one side, and he economic alua ion o he applied o e-
cas on he o he side.” The necessi y o hose b oade conside a ions is highligh ed
by Kmiecik and Zangana (2022), who a gue ha “while compa ing he ex-pos
e o s o di e en o ecas s is a me hod o de e mining which is supe io , i can-
no alone jus i y he alue i adds, i is a he a ela i e measu emen .” The e o e,
he au ho s emphasise he necessi y o conside ing he added business alue esul -
ing om any imp o emen o demand p edic ions, which hey desc ibe as “ o ecas
alue added.” Wi hin his assessmen , hey acknowledge a ious expenses ela ed
o model deploymen , including inancial, empo al and human esou ces (Kmiecik
and Zangana 2022). While he p edominan ocus in he e iewed s udies is on quan-
i a i e s a is ical e o me ics, some s udies a e iden i ied ha o e such a mo e
holis ic assessmen , in eg a ing b oade economic and business dimensions o a
comp ehensi e model assessmen . Fo example, Glaese e al. (2019) look beyond
adi ional e o me ics and e alua e hei models based on i s po en ial inancial
impac on he business. Fo his eason, hey un simula ions o compa e he po en-
ial e enue impac s o using hei algo i hm o op imise e ail loca ions. Simila ly,
Khandani e al. (2010) explo e he p ac ical implica ions o enhancing hei c edi
isk p edic ion model, expec ing a 6 o 25% educ ion in o al losses by iden i y-
ing high- isk demande s ea lie . In addi ion, Miloše ić e al. (2017) in es iga e he
po en ial p o i gain associa ed wi h hei ML chu n p e en ion app oach, which
2791
The applica ion o machine lea ning o demand p edic ion…
p omises o inc ease p o i wo old by p o iding a mo e accu a e iden i ica ion o
demande s likely o chu n.
An essen ial inal s ep in he pos p ocessing phase a ises om he gene al aca-
demic need o iden i y and unde s and he limi a ions associa ed wi h ML p edic-
ions o demand beha iou unde ola ile mac oeconomic condi ions. In o de o
p o ide a concise sys ema isa ion o hese issues and es ic ions, hey a e ex ac ed
om he 64 e iewed s udies and sys ema ically o ganised in o ou main a eas o
limi a ions: (1) da a, (2) gene alisabili y, (3) in e p e abili y and (4) me hodological
and applica ion limi a ions.
An impo an numbe o he iden i ied limi a ions ela e o he (1) da a dimen-
sion, iden i ied as a ele an cons ain in a majo i y o he s udies. These limi a ions
include da a inconsis encies such as missing da a, omi ed a iables o unbalanced
da a due o he o igin o he da ase o he way o da a collec ion, whe e e en small
de ia ions in da a quali y can signi ican ly a ec model ou comes, as se e al au ho s
in he sample indica e (Ghoddusi e al. 2019; Liu e al. 2021). Mo eo e , his ca -
ego y encompasses issues ela ed o a limi ed sample size (e.g., Anand and Mish a
2022; Esmeli e al. 2021; Lin e al. 2022) o a es ic ed numbe o inpu a iables
(e.g., Khandani e al. 2010; Kozak e al. 2021; Raizada and Saini 2021), which can
bo h a ec he obus ness and p edic abili y o a model. Pa icula ly in his esea ch
con ex , in ol ing a apidly changing mac oeconomic en i onmen whe e da a may
no ully e lec e ol ing demand s uc u es, a egula ly ci ed conce n is he use o
ML wi h da a om na owly limi ed ime pe iods o elying excessi ely on his o i-
cal demand da a, which can bu den he an icipa ion o mo e ecen u bulen e en s.
As Cas illo e al. (2017) emphasise “ he e iciency o hese echniques s ongly
depends on he ield o applica ion and he co ec ness o he p oblem da a.” The e-
o e, in ola ile pe iods, a po en ial equi emen esul ing om his limi a ion is he
necessi y o inco po a e su icien demand da a om such ans o ma i e o u bu-
len phases in o model aining o p o ide an accu a e da a ounda ion o gene a -
ing eliable demand p edic ions abou he u u e de elopmen in hese imes. This
is add essed, o example, by Tudo (2022), who speci ically spli s he da a, he eby
allowing o inco po a e da a om he pandemic in o he aining da ase s, enabling
he models o adap o he ola ile pe iod.
The nex ca ego y, (2) gene alisabili y, e e s o limi a ions in alida ing he p o-
posed models o a b oade spec um o demand p edic ion scena ios, esea ch p ob-
lems o domain a eas ha may a ise om a e y speci ic esea ch pe spec i e in
he s udy (Kozak e al. 2021). Fo example, se e al au ho s limi hei s udy esul s
o a speci ic indus y, such as e ail (e.g., Chashmi e al. 2021; Glaese e al. 2019;
Ma ínez-de-Albéniz e al. 2020). O he s limi hei ocus o a pa icula coun y
(e.g., Cas illo e al. 2017; Menhaj and Ka oosi-Kalashami 2022; Wu e al. 2022),
a p edic ion i em (e.g., Punia and Shanka 2022) o a pa icula company (Bi e al.
2022; Raizada and Saini 2021). A consequence o his is he gene al necessi y o
e alua e he b oade applicabili y o speci ic esea ch esul s in ela ion o he p e-
ailing amewo k and he p ac ical ma ke ing p edic ion scena io. Fu he mo e, in
he ML con ex , he scope o gene alisabili y also includes he ac ha no all exis -
ing algo i hms—especially hose ha a e pa icula ly complex o ecen ly de el-
oped—can be aken in o accoun in he modelling o p ep ocessing s ages. Fac o s
2792
M.Mu h e al.
con ibu ing o he limi ed numbe o ML models conside ed include no only he
necessi y o ex ensi e compu ing ime and esou ces (Chen e al. 2021) bu also
he equen eme gence o new models in his ield. This is e idenced by he li -
e a u e sample, as se e al s udies in oduce new models, such as he PLR-ALSTM-
NN (Poly-linea eg ession wi h Long Sho -Te m Memo y) by Ahmed e al. (2022),
he ATLAS (Ad anced Tempo al La en - ac o App oach o Sales o ecas ing) by
Bi e al. (2022) o he EHTS model (Ensemble o Boos ed Hyb id o Deep Lea n-
ing Models and Technical Analysis o Fo ecas ing S ock P ices) by Kama a e al.
(2022).
A u he ca ego y add esses he limi ed (3) in e p e abili y ega ding he
model o he ela ionships i es ablishes, which is add essed in a smalle bu el-
e an numbe o s udies. Consequen ly, p edic ing demand beha iou in a ola ile
mac oeconomic en i onmen in ol es looking no only a he p edic i e capabili-
ies o ML bu also a he equi emen s o s akeholde s in e ms o i s unde -
s andabili y, explainabili y and anspa ency. As Gü ses-T an and Mon i (2022)
obse e ha “mo e complex models eme ge ha a e less anspa en o unde -
s andable o he decision make s. Howe e […] when o ecas models lack in e -
p e abili y […], he us and con idence o s akeholde s can su e when mak-
ing decisions.” Kozak e al. (2021) echo his by s a ing: “I decision-make s a e
no able o in e p e da a p ope ly o a e no able o p epa e decisions, big da a
analy ics p o ides only li le alue.” F om a me hodological pe spec i e, speci ic
app oaches o gain insigh s in o he demand model can be ound in he sample,
like ea u e impo ance assessmen , which es ima es he con ibu ions o indi id-
ual ea u es o imp o e he p edic i e model (Ma and Fildes 2020). Ano he ech-
nique is sensi i i y analysis, whe e he changes in he ou pu model a e obse ed
when speci ic inpu s a e adjus ed (Bohanec e al. 2017). To his end, Bohanec
e al. (2017) wa n ha “ he explana ions closely ollow he p edic ion model; i
he model is w ong o pe o ms poo ly, he explana ions will e lec ha .” S udy
examples o such e alua ions can be ound in Cla e ia e al. (2020), who ana-
lyse he impac o a iables on hei GDP g ow h p edic ion model and ind ha
unde es ima ed ye c ucial su ey a iables a e cu en ly no included in o icial
s a is ics. Ano he example can be ound in Chen e al. (2021), who use ML in
pu chasing beha iou esea ch and model cus ome p e e ences when conside ing
he in luence o sales p ices. A e modelling, he au ho s ex ac he p ice sensi-
i i y o ce ain ca ego ies and ind ha p oduc s hey see as essen ial o li e, like
eggs, milk and b ead, indica e a low p ice sensi i i y in he model. In con as ,
ca ego ies such as canned ood o eady- o-ea meals, which ha e a longe shel
li e and can be easily s o ed, show a highe model p ice sensi i i y. They claim
ha hese insigh s can guide ailo ed p omo ions based on consume p ice beha -
iou a he ca ego y o p oduc le el.
The inal limi a ions encoun e ed in he ansi ion om s udy esul s o (4)
eal-wo ld applica ions as well as no able me hodological limi a ions. Rega d-
ing he i s aspec , he con ibu ion o Balles a e al. (2019) om he li e a-
u e sample, wi h he acknowledgemen o he s udy o Lambe on and S ephen
(2016), can be used o help cha ac e ise his p oblem. They c i ically discuss he
“myopic” app oach common in scien i ic analyses, which o en ends o na ow
2793
The applica ion o machine lea ning o demand p edic ion…
he ac ual scope o a much la ge and mul i-laye ed eal-wo ld p oblem o jus a
ew ace s. This can esul , o example, om he inabili y o ce ain models o
include all ele an ac o s ha migh in luence ou comes in eal-wo ld scena ios
o om he di icul y o measu ing hese e ec s— anging om b oade ex e nal
a iables such as ce ain p e ailing mac oeconomic condi ions o mo e de ailed
in o ma ion on consume s’ demand beha iou (e.g., Bi e al. 2022; Khandani
e al. 2010; Hasheminejad e al. 2022). Howe e , Anand and Mish a (2022) also
emphasise he inhe en complexi y o eal-wo ld applica ions and, in pa icula ,
he in icacies ha a ise om human beha iou , desc ibing i as “nonlinea , non-
pa ame ic, i a ional, and ime- a ian .” Fu he mo e, he necessi y o con inued
p ac ical e i ica ion is exp essed as a conce n ega ding eal-wo ld ope a ional
con ex s. As o mula ed by Kozak e al. (2021) “mos o he decision-making
echniques use de e minis ic machine lea ning (ML) echniques bu un o una ely,
hey do no ake in o accoun he a ie y and ola ili y o decision-making si u-
a ions and do no allow o a mo e lexible app oach, i.e., adjus ed o changing
en i onmen al condi ions o changing managemen p io i ies.” Unexpec ed mac-
oeconomic ola ili y can exace ba e his challenge, as, o example, he de elop-
men o ce ain demand in luencing a iables migh no be an icipa ed in ad ance
a he ime when he demand p edic ion is gene a ed (Cas illo e al. 2017). Apa
om he issue o eal-wo ld applica ions, he second aspec o be add essed is
ha o me hodological limi a ions, which co e a wide a ie y o speci ic p ob-
lems. These can be, o example, me hodological sho comings on a echnical
le el ega ding he way he model and i s pa ame e s a e gene a ed in he speci ic
con ex o ML—which can be he p oblem o o e i ing he model, he pa ly
manual sea ch o model se ings (hype pa ame e s) o he o e all s ong eliance
on he p inciple o ial-and-e o . All o hese can signi ican ly complica e he
ask o placing he ML model and he ela ionships i con ains on a solid heo e i-
cal ounda ion (Cla e ia e al. 2020; Ghoddusi e al. 2019). Co eney e al. (2016)
unde sco e his, s a ing: “We a gue ha i is i al o use heo y as a guide o
expe imen al design o maximal e iciency o da a collec ion and o p oduce eli-
able p edic i e models and concep ual knowledge,” and Bi e al. (2022) conclude:
“Inco po a ing economic heo y conside a ions in o machine lea ning models
can p o ide signi ican addi ional ad an ages, and, hus, cons i u es a p omising
di ec ion o u u e wo k.”
4.4 Limi a ions
Also he esea ch conduc ed in his s udy o de elop he sys ema ic li e a u e
e iew is bound by ce ain limi a ions. Ou lining hem helps o con ex ualise his
s udy’s ou come and also o e s an oppo uni y o build on his ounda ion in u -
he esea ch. Fi s ly, a limi a ion a ises om he selec ion o he Web o Science-
da abase o his li e a u e e iew, based on he esea ch s udy o Gusenbaue and
Haddaway (2019), as i s scope is es ic ed o he jou nals indexed he ein. This
implies ha o ma s ou side jou nal publica ions, such as con e ence p oceedings
o published books (Kepes e al. 2012; Vlačić e al. 2021), as well as non-academic
2794
M.Mu h e al.
jou nals o g ey li e a u e (Rejeb e al. 2020), a e no speci ically co e ed. The e-
o e, as he li e a u e sample is ocused on jou nal publica ions o ensu e compa a-
bili y o he da a h ough s ic jou nal e iew p ocesses, he e is a isk o exclud-
ing some cu en esea ch indings. Fu he mo e, Ghoddusi e al. (2019) highligh
an una oidable bu ne e heless no able es ic ion o in e disciplina y esea ch
in ol ing compu e science by poin ing o unpublished de elopmen s o ML algo-
i hms ha a e only used o comme cial business pu poses and a e he e o e no
accessible ia s udies. In gene al, howe e , he in e disciplina y esea ch o ien a ion
in his li e a u e e iew allows o he o ma ion o an o e iew ac oss disciplines
ha goes beyond he mains eam o a pa icula ield and ha p o ides a di e en i-
a ed and mo e holis ic new pe spec i e (Ghoddusi e al. 2019; Rejeb e al. 2020).
As a second limi a ion o his pape , i should be men ioned ha he p ecisely
de ined sea ch s ing no only in ol es he isk o alse posi i es, hus including non-
ele an s udies (Linnenluecke e al. 2019; Xiao and Wa son 2019), bu also a isk o
alse nega i es h ough excluding pape s ha all ou side he sea ch s ing bu ha
may p o ide ele an knowledge (Du ugbo and Al-Balushi 2022). Ano he conse-
quence o he sea ch s ing used is ha he iden i ied s udies la gely ocus on spe-
ci ic indus ies such as consume goods and e ail (n = 22) as well as economics and
inancial managemen (n = 13). Also, he inclusion o only English language s udies
could igno e ele an indings om non-English-language s udies, as he analysis
in his pape highligh s ele an geog aphical esea ch cen es, o example in Asia
(n = 100) o in la ge pa s o Eu ope (n = 86), ha a e non-na i e English speaking.
Thi d, he de ini ion o he inclusion and exclusion c i e ia s ill allows esea che s
a ce ain leeway, which can a ec in e a e eliabili y (Cla k e al. 2021; Xiao and
Wa son 2019). Howe e , he e iew p o ocol wi h an objec i e lis o c i e ia clea ly
limi s indi idual subjec i i y and hus ensu es he gene al eplicabili y o he p e i-
ously p esen ed esea ch indings (Vlačić e al. 2021).
5 Conclusion
This sys ema ic li e a u e e iew summa ises ecen esea ch e idence on he e ec-
i e applica ion o ML in imes o mac oeconomic ola ili y, pa icula ly ocusing
on demand p edic ion. The need o syn hesise his speci ic esea ch knowledge is
mo i a ed by he cu en pe iod, ma ked by p onounced global unce ain y and
u bulence (see Du s e al. 2022; Du ugbo and Al-Balushi 2022). He e, accu a e
demand p edic ion is c ucial o o ganisa ional unc ions like ma ke ing (Kozak
e al. 2021; Seyedan and Ma akhe i 2020). The e o e, his hema ic a ea is assessed
by analysing a sample o 64 jou nal a icles om an ini ial pool o 2877 pape s,
which we e selec ed using a ca alogue o c i e ia speci ically de eloped as pa o
he e iew p o ocol (Xiao and Wa son 2019). Using an in eg a i e app oach wi h a
hyb id me hodology (T app 2012), ele an knowledge is syn hesised om he inal
sample (Sec .4), he eby con ibu ing o closing he gap iden i ied in p e ious li -
e a u e e iews. This sys ema ic li e a u e e iew hus con ibu es o he ongoing
academic discou se on he e ec i e use o ML in applica ion p ac ice, in eg a ing
2795
The applica ion o machine lea ning o demand p edic ion…
mul idisciplina y insigh s om a ious s eams wi hin compu e science and man-
agemen science.
A he beginning o his sys ema ic li e a u e e iew, an analysis o he cha ac-
e is ics and abs ac s o he a icles (Sec s. 4.1 and 4.2) e eals ha li e a u e since
2010 app oxima es an exponen ial end cu e in publica ions, which demons a es
e idence o a g owing scien i ic in e es in his in e disciplina y esea ch opic. A
geog aphical concen a ion o academic discussion is ound, wi h Asia accoun -
ing o he la ges sha e o publica ions, ollowed by Eu ope and No h Ame ica,
wi h he U.S. and China as he leading coun ies. The jou nal “Expe Sys ems wi h
Applica ions” and au ho s such as Hyndman and B eiman dese e special men ion,
as hey a e iden i ied as essen ial e e ences in his complex ield.
In he u he cou se, his e iew pape also p o ides a mo e de ailed pe spec-
i e wi h a syn hesis o he ull ex s, ch onologically ou lining he ML appli-
ca ion p ocess (Sec .4.3). The analyses in he p ep ocessing phase (Sec .4.3.1)
indica e ha o e 75% o he empi ical s udies in ol e he use o mul i a ia e
models, which emphasises he need o conside mul iple ac o s o cap u e he
mul i ac o ial na u e o demand beha iou in a ola ile en i onmen . Addi ion-
ally, o wa d-looking a iables a e iden i ied as impo an inpu a iables o
demand p edic ions, o e ing ea ly signals o mac oeconomic ends and u ning
poin s. Examples ci ed include ma ke pa icipan s’ expec a ions and sen imen s
ia economic indica o s o sea ch ends. In addi ion, he esea ch ega ding he
modelling phase (Sec .4.3.2) ini ially e eals ha a majo i y o he s udies illus-
a e an ou pe o mance o ML o e o he p edic ion me hods (e.g., Ma and Fildes
2020; Alb ech e al. 2021). This implies he gene al sui abili y o ML o model-
ling complex demand pa e ns and in e dependencies in dynamic ma ke condi-
ions. Howe e , he esea ch also iden i ies s udies whe e al e na i e quan i a i e
models su pass ML (e.g., Abolghasemi e al. 2020; Lo i e al. 2023), indica ing
ha ML models do no possess uni e sal supe io i y in his con ex . Fu he mo e,
he analysis deli e s an assessmen o he usage equency o di e en models in
he empi ical s udies, as well as he equency wi h which models a e explici ly
ecommended by esea che s due o hei supe io p edic i e pe o mance. The e,
i becomes e iden ha supe ised ML algo i hms a e p e alen in app oxima ely
90% o he s udies. Fu he mo e, a ound hal o he esea ch pape s demons a e
he ou pe o mance o combining mul iple models ia hyb id o ensemble model-
ling, unde sco ing hei signi icance o esilien demand p edic ion in complex
da a en i onmen s. The analysis also highligh s speci ic ML algo i hms, ecom-
mending hei inclusion o modelling in his con ex . A ound 40% o he s ud-
ies epo ha A i icial Neu al Ne wo ks, ei he ully o pa ially, con ibu e o
he bes -pe o ming model, wi h Recu en Neu al Ne wo ks pa icula ly s anding
ou o hei abili y o cap u e ime-dependen and non-linea ela ionships. The
analysis iden i ies u he key ML algo i hms in his con ex —speci ically Ran-
dom Fo es s/Decision T ees, and G adien Boos ing—as pa icula ly e ec i e,
each demons a ing supe io pe o mance in o e 10% o he s udies. Addi ional
modelling s a egies o accoun o demand ola ili y a e also discussed, such as
employing he Coe icien o Va ia ion, demand decomposi ion, pooling, o spe-
ci ic ime windows. Finally, in he pos p ocessing phase (Sec .4.3.3), e alua ion
2796
M.Mu h e al.
s a egies a e iden i ied ha accoun no only o s a is ical pe o mance bu
also o he b oade business-o ien ed implica ions o ML demand p edic ions.
Mo eo e , om he 64 s udies, ou main ypes o limi a ions a e syn hesised o
conside when implemen ing ML demand p edic ion unde mac oeconomic ola-
ili y: (1) da a, (2) gene alisabili y, (3) in e p e abili y, as well as (4) me hodo-
logical and applica ion- ela ed limi a ions.
O e all, based on he esea ch indings, an e idence-based guideline can be p o-
ided o ein o ce business a eas such as ma ke ing unde he p e ailing condi ions.
Fu u e esea ch di ec ions can be de i ed om he concep ualisa ion o he ield as
p esen ed in his sys ema ic li e a u e e iew (Sec .4.2). As de e mined in his pape
by a co espondence analysis o he ex ac ed abs ac e ms, he scien i ic ield can
be ca ego ised in o i e ocus esea ch a eas: (F1) ML me hods and hei applica ion
con ex , (F2) Economic ac o s, (F3) Time se ies o ecas ing, (F4) Neu al ne wo k
p edic ions, and (F5) Cus ome amewo ks. Wi hin hese a eas, u he in-dep h
in es iga ion is encou aged. Mo eo e , addi ional concep ual wo k is impo an ,
such as de eloping a s uc u ed amewo k o p o ide s a e-o - he-a di ec i es o
applying ML in demand p edic ion (Alsha e e al. 2022; Du ugbo and Al-Balushi
2022; Ghoddusi e al. 2019). Resea ch oppo uni ies also s em om expanding upon
he key insigh s h ough u he empi ical in es iga ions.
In a b oade sense, his pape encou ages he os e ing o mo e in eg a ed
app oaches ha ha monise igo ous algo i hmic me hods wi h p ac ical business
objec i es. I emphasises he need o a pa icula ly holis ic pe spec i e o conside
in e disciplina y in e ac ions be ween esea ch a eas, as seen, o example, espe-
cially be ween he disciplines o compu e science and managemen science. The
comp ehensi e o e iew esul ing om his s udy is in ended o p o ide a pla o m
o u u e de elopmen s in his apidly e ol ing ield. By in oducing a new iew and
a deepe sys ema isa ion o he esea ch a ea, his sys ema ic li e a u e e iew can
ac as a obus scien i ic ounda ion o u he academic and p ac ical endea ou s.
Supplemen a y In o ma ion The online e sion con ains supplemen a y ma e ial a ailable a h ps:// doi.
o g/ 10. 1007/ s11301- 024- 00447-8.
Funding Open Access unding enabled and o ganized by P ojek DEAL. The au ho s decla e ha no
unds, g an s, o o he suppo we e ecei ed du ing he p epa a ion o his manusc ip .
Da a a ailabili y As a suppo ing documen , an Excel ile wi h a de ailed o e iew o all pape s inco -
po a ed in he li e a u e sample is p o ided wi h he manusc ip submission. This ile complemen s he
in o ma ion p o ided in his sys ema ic li e a u e e iew, which al eady includes speci ics ega ding he
da a collec ion, such as he li e a u e sou ces, sea ch s ing, and selec ion c i e ia. Fo addi ional analyses
and in o ma ion ega ding he esea ch pape , he co esponding au ho can be con ac ed. We ensu e ha
all da a suppo ing his s udy’s esul s a e a ailable in he a icle, supplemen a y documen s, o h ough
he co esponding au ho , upholding anspa ency and ull accessibili y o he scien i ic communi y.
Decla a ions
Compe ing in e es s The au ho s ha e no ele an inancial o non- inancial in e es s o disclose.
Open Access This a icle is licensed unde a C ea i e Commons A ibu ion 4.0 In e na ional License,
which pe mi s use, sha ing, adap a ion, dis ibu ion and ep oduc ion in any medium o o ma , as long
as 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
2797
The applica ion o machine lea ning o demand p edic ion…
Commons licence, and indica e i changes we e made. The images o o he hi d pa y ma e ial in his
a icle a e included in he a icle’s C ea i e Commons licence, unless indica ed o he wise in a c edi line
o he ma e ial. I ma e ial is no included in he a icle’s C ea i e Commons licence and you in ended
use is no pe mi ed by s a u o y egula ion o exceeds he pe mi ed use, you will need o ob ain pe mis-
sion di ec ly om he copy igh holde . To iew a copy o his licence, isi h p://c ea i ecommons.o g/
licenses/by/4.0/.
Re e ences
Abolghasemi M, Beh E, Ta G, Ge lach R (2020) Demand o ecas ing in supply chain: he impac o
demand ola ili y in he p esence o p omo ion. Compu Ind Eng 142:106380. h ps:// doi. o g/ 10.
1016/j. cie. 2020. 106380
Ahmed S, Chak abo y RK, Essam DL, Ding W (2022) Poly-linea eg ession wi h augmen ed long
sho e m memo y neu al ne wo k: p edic ing ime se ies da a. In Sci 606:573–600. h ps:// doi.
o g/ 10. 1016/j. ins. 2022. 05. 078
Alb ech T, Rausch TM, De a ND (2021) Call me maybe: me hods and p ac ical implemen a ion o
a i icial in elligence in call cen e a i als’ o ecas ing. J Bus Res 123:267–278. h ps:// doi. o g/ 10.
1016/j. jbus es. 2020. 09. 033
Alsha e A, Agga wal K, Sonia KM, Mish a A (2022) Re iew o ML and Au oML solu ions o o e-
cas ime-se ies da a. A ch Compu Me hod Eng 29:5297–5311. h ps:// doi. o g/ 10. 1007/
s11831- 022- 09765-0
Anand S, Mish a K (2022) Iden i ying po en ial millennial cus ome s o inancial ins i u ions using
SVM. J Financ Se Ma k 27:335–345. h ps:// doi. o g/ 10. 1057/ s41264- 021- 00128-7
Angus RW, Packa d MD, Cla k BB (2023) Dis inguishing unp edic abili y om unce ain y in en ep e-
neu ial ac ion heo y. Small Bus Econ 60:1147–1169. h ps:// doi. o g/ 10. 1007/ s11187- 022- 00651-4
A ia M, Cuccu ullo C (2023) Package ’bibliome ix’–comp ehensi e science mapping analysis. CRAN.
h ps:// c an. - p oje c . o g/ web/ packa ges/ bibli ome ix/ bibli ome ix. pd
A ia M, Cuccu ullo C (2017) Bibliome ix: an R- ool o comp ehensi e science mapping analysis. J
In o me 11:959–975. h ps:// doi. o g/ 10. 1016/j. joi. 2017. 08. 007
A un aj NS, Ah ens D (2015) A hyb id seasonal au o eg essi e in eg a ed mo ing a e age and quan ile
eg ession o daily ood sales o ecas ing. In J P od Econ 170:321–335. h ps:// doi. o g/ 10. 1016/j.
ijpe. 2015. 09. 039
Balles a MT, G au-Ca les P, Sáinz J (2019) P edic ing cus ome quali y in e-comme ce social ne -
wo ks: a machine lea ning app oach. Re Manag Sci 13:589–603. h ps:// doi. o g/ 10. 1007/
s11846- 018- 0316-x
Bi X, Adoma ičius G, Li W, Qu A (2022) Imp o ing sales o ecas ing accu acy: a enso ac o iza ion
app oach wi h demand awa eness. INFORMS J Compu 34:1644–1660. h ps:// doi. o g/ 10. 1287/
ijoc. 2021. 1147
Black JE, Kuepe JK, Williamson TS (2022) An in oduc ion o machine lea ning o classi ica ion and
p edic ion. Fam P ac 40:200–204. h ps:// doi. o g/ 10. 1093/ amp a/ cmac1 04
Bohanec M, Bo š na MK, Robnik-Šikonja M (2017) Explaining machine lea ning models in sales p e-
dic ions. Expe Sys Appl 71:416–428. h ps:// doi. o g/ 10. 1016/j. eswa. 2016. 11. 010
Boo h A, Su on A, Papaioannou D (2016) Sys ema ic app oaches o a success ul li e a u e e iew, 2nd
edn. Sage, Los Angeles
B eiman L (2001) Random o es s. Mach Lea n 45:5–32. h ps:// doi. o g/ 10. 1023/A: 10109 33404 324
Bukha i AH, Raja MAZ, Sulaiman M, Islam S, Shoaib M, Kumam P (2020) F ac ional neu o-sequen ial
ARFIMA-LSTM o inancial ma ke o ecas ing. IEEE Access 8:71326–71338. h ps:// doi. o g/ 10.
1109/ access. 2020. 29857 63
Ca iolle J, Goujon M (2013) Measu ing mac oeconomic ins abili y: a c i ical su ey illus a ed wi h
expo s se ies. J Econ Su 29:1–26. h ps:// doi. o g/ 10. 1111/ joes. 12036
Cas illo PA, Mo a AM, Fa is H, Me elo JJ, Ga cia-Sanchez P, Fe nandez-A es AJ, De las Cue as P,
Ga cia-A enas MI (2017) Applying compu a ional in elligence me hods o p edic ing he sales o
newly published books in a eal edi o ial business managemen en i onmen . Knowl-Based Sys
115:133–151. h ps:// doi. o g/ 10. 1016/j. knosys. 2016. 10. 019
2798
M.Mu h e al.
Chashmi AJ, Rahma i V, Rezaso oush B, Alamo i MM, Aska i M, Khalili FH (2021) P edic ing cus ome
u no e using ecu si e neu al ne wo ks. Wi el Commun Mob Compu 2021:1–11. h ps:// doi. o g/
10. 1155/ 2021/ 66230 52
Chen M (2022) A MapReduce-based a i icial neu al ne wo k chu n p edic ion o music s eaming se -
ice. In J Compu Sci Ne w Secu 22:55–60. h ps:// doi. o g/ 10. 22937/ IJCSNS. 2022. 22.1.9
Chen SS, Choubey B, Singh V (2021) A neu al ne wo k based p ice sensi i e ecommende model o
p edic cus ome choices based on p ice e ec . J Re ail Consum Se 61:102573. h ps:// doi. o g/
10. 1016/j. j e c onse . 2021. 102573
Cla k WR, Cla k LA, Ra o DM, Williams RI (2021) Ex ending Fisch and Block’s (2018) ips o a sys-
ema ic e iew in managemen and business li e a u e. Manag Re Q 71(1):215–231. h ps:// doi.
o g/ 10. 1007/ s11301- 020- 00184-8
Cla e ia O, Mon e E, To a S (2020) Economic o ecas ing wi h e ol ed con idence indica o s. Econ
Model 93:576–585. h ps:// doi. o g/ 10. 1016/j. econm od. 2020. 09. 015
Con ad CA (2020) Angewand e mak oökonomie—eine p axisbezogene ein üh ung. Sp inge , Wies-
baden. h ps:// doi. o g/ 10. 1007/ 978-3- 658- 30055-5
Con e as-Masse R, Ochoa-Zezza i A, Pe ez-Dominguez L, Naudasche K (2022) Da a-d i en en e -
p ises o economic eco e y a e pandemic: a s udy case in consume -packaged goods indus y.
In J Comb Op im P obl In o m 13:33–39
Coussemen K, Van den Poel D (2008) Chu n p edic ion in subsc ip ion se ices: an applica ion o
suppo ec o machines while compa ing wo pa ame e -selec ion echniques. Expe Sys Appl
34:313–327. h ps:// doi. o g/ 10. 1016/j. eswa. 2006. 09. 038
Co eney PV, Doughe y ER, High ield RR (2016) Big da a need big heo y oo. Phil T ans R Soc A
374:20160153. h ps:// doi. o g/ 10. 1098/ s a. 2016. 0153
Danahe PJ, B odie RJ (2000) Unde s anding he cha ac e is ics o p ice elas ici ies o equen ly pu -
chased packaged goods. J Ma k Manag 16:917–936. h ps:// doi. o g/ 10. 1362/ 02672 57007 84683 726
Denye D, T an ield D (2009) P oducing a sys ema ic e iew. In: Buchanan DA, B yman A (eds) The
sage handbook o o ganiza ional esea ch me hods. Sage Publica ions L d, London, pp 671–689
Denye D, T an ield D, Van Aken JE (2008) De eloping design p oposi ions h ough esea ch syn hesis.
O gan S ud 29:393–413. h ps:// doi. o g/ 10. 1177/ 01708 40607 088020
Don hu N, Kuma S, Mukhe jee D, Pandey N, Lim WM (2021) How o conduc a bibliome ic analysis:
an o e iew and guidelines. J Bus Res 133:285–296. h ps:// doi. o g/ 10. 1016/j. jbus es. 2021. 04. 070
Dö ne A, Maisch M, Münch a h J (2023) Beben an den mä k en. Handelsbla , Düsseldo , pp 6–8
Duca JV, Muellbaue J, Mu phy A (2010) Housing ma ke s and he inancial c isis o 2007–2009: lessons
o he u u e. J Financ S abil 6:203–217. h ps:// doi. o g/ 10. 1016/j. j s. 2010. 05. 002
Duncan GT, Go WL, Szczypula J (2001) Fo ecas ing analogous ime se ies. In: A ms ong JS (ed) P in-
ciples o o ecas ing: a handbook o esea che s and p ac i ione s. Sp inge , Bos on, pp 195–213.
h ps:// doi. o g/ 10. 1007/ 978-0- 306- 47630-3_ 10
Du s S, Foli S, Ed a dsson IR (2022) A sys ema ic li e a u e e iew on knowledge managemen in
SMEs: cu en ends and u u e di ec ions. Manag Re Q 74:263–288. h ps:// doi. o g/ 10. 1007/
s11301- 022- 00299-0
Du ugbo CM, Al-Balushi Z (2022) Supply chain managemen in imes o c isis: a sys ema ic e iew.
Manag Re Q 73:1179–1235. h ps:// doi. o g/ 10. 1007/ s11301- 022- 00272-x
E a MIA, Hájek P, Abedin MZ, Azad RU, Jabe MA, Adi ya S, Hassan MK (2022) Deep-lea ning model
using hyb id adap i e end es ima ed se ies o modelling and o ecas ing sales. Ann Ope Res
22:1–32. h ps:// doi. o g/ 10. 1007/ s10479- 022- 04838-6
El Shoubaki A, Block J, Lasch F (2021) The couple business as a unique o m o business: a e iew o
he empi ical e idence. Manag Re Q 72:115–147. h ps:// doi. o g/ 10. 1007/ s11301- 020- 00206-5
E ns J, Schmid J, Beneken G (2020) G undku s in o ma ik: g undlagen und konzep e ü die e olg e-
iche IT-p axis eine um assende, p axiso ien ie e ein üh ung, 7 h edn. Sp inge , Wiesbaden
Esmaeilzadeh P (2023) The ole o Cha GPT in dis up ing concep s, changing alues, and chal-
lenging e hical no ms: a quali a i e s udy. AI E hics 2023:1–14. h ps:// doi. o g/ 10. 1007/
s43681- 023- 00338-w
Esmeli R, Bade -El-Den M, Abdullahi H (2021) Towa ds ea ly pu chase in en ion p edic ion in
online session based e ailing sys ems. Elec on Ma k 31:697–715. h ps:// doi. o g/ 10. 1007/
s12525- 020- 00448-x
Eu opean Cen al Bank (2023) Eu osys em s a mac oeconomic p ojec ions o he eu o a ea-Decembe
2023. h ps:// www. ecb. eu opa. eu/ pub/ pd / o he / ecb. p oje c ion s2023 12_ eu os ys em s a ~9a39a b5088.
en. pd
