Sa ode, Mohi S.; Kuma , Anil; P asad, Abhiji ; She y, Abhishek
A icle
Enhancing p icing s a egies in he a e ma ke sec o wi h
machine lea ning
Mode n Supply Chain Resea ch and Applica ions
P o ided in Coope a ion wi h:
Eme ald Publishing Limi ed
Sugges ed Ci a ion: Sa ode, Mohi S.; Kuma , Anil; P asad, Abhiji ; She y, Abhishek (2024) :
Enhancing p icing s a egies in he a e ma ke sec o wi h machine lea ning, Mode n Supply Chain
Resea ch and Applica ions, ISSN 2631-3871, Eme ald, Bingley, Vol. 6, Iss. 4, pp. 411-423,
h ps://doi.o g/10.1108/MSCRA-10-2023-0042
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/314934
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/
Enhancing p icing s a egies in he
a e ma ke sec o wi h
machine lea ning
Mohi S. Sa ode, Anil Kuma , Abhiji P asad and Abhishek She y
Daimle T uck Inno a ion Cen e India, Bangalo e, India
Abs ac
Pu pose –This esea ch explo es he applica ion o machine lea ning o op imize p icing s a egies in he
a e ma ke sec o , pa icula ly ocusing on pa s wi h no assigned alues and he de ec ion o ou lie s. The s udy
emphasizes he need o inco po a e echnical ea u es o imp o e p icing accu acy and decision-making.
Design/me hodology/app oach –The me hodology in ol es da a collec ion om web sc aping and backend
sou ces, ollowed by da a p ep ocessing, ea u e enginee ing and model selec ion o cap u e he echnical
a ibu es o pa s. A Random Fo es Reg esso model is chosen and ained o p edic p ices, achie ing a 76.14%
accu acy a e.
Findings –The model demons a es accu a e p ice p edic ion o pa s wi h no assigned alues while emaining
wi hin an accep able p ice ange. Addi ionally, ou lie s ep esen ing ex eme p icing scena ios a e success ully
iden i ied and p edic ed wi hin he accep able ange.
O iginali y/ alue –This esea ch b idges he gap be ween indus y p ac ice and academic esea ch by
demons a ing he e ec i eness o machine lea ning o a e ma ke p icing op imiza ion. I o e s an app oach
o add ess he challenges o p icing pa s wi hou assigned alues and iden i ying ou lie s, po en ially leading o
inc eased e enue, sha pe p icing ac ics and a compe i i e ad an age o a e ma ke companies.
Keywo ds A e ma ke , Machine lea ning, P ice p edic ion
Pape ype Resea ch pape
1. In oduc ion
In oday’s compe i i e business en i onmen , p icing decisions a e c i ical o an o ganiza ion’s
inancial success (Kalpana e al., 2022). Businesses aiming o compe i i eness and inancial
success mus imp o e hei p icing s a egies. An eme ging ool, machine lea ning-based
p edic i e p icing, uses his o ical da a o quickly o ecas op imal p ices, assis ing in se ing
ai a es and adap ing o changing ma ke condi ions (Bane jee and Bandyopadhyay, 2020).
This app oach employs machine lea ning echniques o deciphe massi e amoun s o da a,
e ealing in ica e pa e ns and de e mining he mos p o i able p ice poin s based on cus ome
p e e ences, ma ke dynamics, and co po a e goals (Gup a and Pa hak, 2014;Man ala e al.,
2006). This ans o ma i e me hodology enables da a-d i en p icing decisions, g an ing
o ganiza ions a compe i i e edge in p icing s a egies.
Se e al s udies ha e explo ed he applica ion o machine lea ning o p icing op imiza ion
in a ious ields. Fo example, adi ional me hods we e compa ed agains machine lea ning
me hods such as Random Fo es , G adien Boos ed Machines, and Deep Lea ne s in he
insu ance indus y, highligh ing he e ec i eness o G adien Boos ing Me hods (Spedica o
e al., 2018). Ou esea ch builds on his by applying a Random Fo es Reg esso model in he
a e ma ke sec o .
Dynamic p icing in e-comme ce has been explo ed using G adien Boos ing Machines
(GBMs), showing supe io pe o mance in cap u ing complex non-linea p icing pa e ns
Mode n Supply
Chain Resea ch
and Applica ions
411
© Mohi S. Sa ode, Anil Kuma , Abhiji P asad and Abhishek She y. Published in Mode n Supply Chain
Resea ch and Applica ions. Published by Eme ald Publishing Limi ed. This a icle is published unde he
C ea i e Commons A ibu ion (CC BY 4.0) licence. Anyone may ep oduce, dis ibu e, ansla e and
c ea e de i a i e wo ks o his a icle ( o bo h comme cial and non-comme cial pu poses), subjec o ull
a ibu ion o he o iginal publica ion and au ho s. The ull e ms o his licence may be seen a h p://
c ea i ecommons.o g/licences/by/4.0/legalcode
The cu en issue and ull ex a chi e o his jou nal is a ailable on Eme ald Insigh a :
h ps://www.eme ald.com/insigh /2631-3871.h m
Recei ed 6 Oc obe 2023
Re ised 24 Janua y 2024
13 June 2024
10 Sep embe 2024
Accep ed 13 Oc obe 2024
Mode n Supply Chain Resea ch and
Applica ions
Vol. 6 No. 4, 2024
pp. 411-423
Eme ald Publishing Limi ed
2631-3871
DOI 10.1108/MSCRA-10-2023-0042
(Youbi e al., 2023). Simila ly, machine lea ning algo i hms like Long Sho -Te m Memo y
Ne wo ks (LSTM), Con olu ional Neu al Ne wo ks (CNN), and Suppo Vec o Reg ession
(SVR) ha e been used o s ock p ice p edic ion, wi h SVR achie ing he highes accu acy
(Chen, 2020). Fu he s udies ha e demons a ed he e ec i eness o Random Fo es s and
A i icial Neu al Ne wo ks (ANN) in p edic ing s ock closing p ices (Vijh e al., 2020).
The use o machine lea ning o p ope y p ice p edic ion has shown ha Random Fo es
and GBM algo i hms pe o m well (Ho e al., 2020). Mo eo e , he applica ion o machine
lea ning o daily commodi y p ice p edic ion has been highligh ed, wi h ANNs showing
e ec i eness bu sugges ing he inco po a ion o domain knowledge and ea u e enginee ing
o imp o emen (Amin, 2020).
Complex p ice p edic ion asks in ol e nume ous ac o s in luencing p ice changes.
T adi ional me hods o en s uggle o accoun o hese complexi ies, while machine lea ning
has shown p omise in op imizing p ices (Indi a e al., 2023). Addi ionally, a model-based
p icing amewo k o machine lea ning models has been p oposed o add ess gaps in da a
ma ke p icing, demons a ing high e enue po en ial and low un ime cos s (Chen
e al., 2019).
Exis ing li e a u e also in es iga es mic o-ma ke ing p icing s a egies based on
supe ma ke scanne da a (Mon gome y, 1997), o ecas s s ock p ices using a ious models
and da a ep esen a ion echniques (Pa el e al., 2015), in es iga es p icing s a egies in B2B
a e ma ke s based on i m size, indus y, and loca ion (Gunaydan, 2023), and op imizes p ices
in dynamic ma ke s wi h limi ed in o ma ion (Dodin e al., 2021).
1.1 Gap in li e a u e
Al hough machine lea ning is widely used in p icing, i is s ill necessa y o apply i in ce ain
indus ies, such as he a e ma ke , which handles eplacemen pa s and componen s o
goods ha ha e al eady been made. The ield o machine lea ning in p icing has been he
subjec o ex ensi e esea ch in he li e a u e. Spedica o e al. (2018) conduc ed a compa ison
be ween machine lea ning models and adi ional p icing echniques. These s udies do no ake
in o accoun he unique ea u es o he a e ma ke sec o , whe e echnical a ibu es o pa s
ha e a signi ican impac on p icing, no do hey concen a e on speci ic me hodological
p ocedu es. This d awback may be seen in esea ch by Youbi e al. (2023), al hough Youbi’s
echnique wo ks well in dynamic con ex s, i is no di ec ly applicable o he a e ma ke sec o
since i igno es echnical ac o s ha ha e a big in luence on p icing decisions. Finding
i egula obse a ions ha poin o mis akes, poo da a quali y, o unusual p icing ends is he
i s s ep in de ec ing anomalies in p icing da a. Simila ly, ou lie de ec ion iden i ies da a
poin s ha di e signi ican ly om he majo i y, equen ly ep esen ing ex eme p ice poin s
o unique ma ke ci cums ances.
Fu he mo e, esea ch by Indi a e al. (2023) emphasizes he impo ance o inco po a ing
indus y-speci ic da a o accu a e p ice p edic ion. I highligh s he need o mo e specialized
app oaches in sec o s wi h unique cha ac e is ics, such as he a e ma ke . Indi a’s wo k poin s
ou ha exis ing models o en ail o accoun o echnical speci ica ions, which a e c i ical in
de e mining p ices o a e ma ke pa s. Ou s udy di ec ly add esses hese gaps by ocusing
on he unique needs o he a e ma ke indus y, whe e he p icing o componen s like uel
anks is de e mined no only by ma ke dynamics bu also by echnical a ibu es such as
ma e ial quali y, size, and du abili y.
By o e ing a ho ough, domain-speci ic echnique o p ice p edic ion and anomaly
iden i ica ion, ou esea ch expands on he undamen al s ages p esen ed by Spedica o e al.
(2018). Ou s udy uses a Random Fo es Reg ession model ained on a da ase supplemen ed
wi h echnical speci ica ions o a e ma ke pa s, in con as o he gene ic echniques in he
li e a u e cu en ly in publica ion. This ocus on echnical elemen s dis inguishes ou esea ch
om ea lie s udies and o e s a mo e accu a e and ele an app oach o he a e ma ke
sec o .
MSCRA
6,4
412
1.2 Resea ch objec i es
Ensu ing cus ome sa is ac ion is c i ical o signi ican playe s in he ucking indus y.
By ensu ing pa s a ailabili y a he igh ime, loca ion, and p ice wi hou sac i icing quali y,
down ime can be educed, and meaning ul business oppo uni ies can be e ained. The
a e ma ke indus y, whe e p icing decisions di ec ly impac e enue and p o i abili y, can
bene i signi ican ly om accu a ely o ecas ing op imal pa s p ices, s iking a balance
be ween a ac ing cus ome s and maximizing p o i ma gins.
The p ima y objec i e o his esea ch is o add ess he challenges o p edic ing p ices o
a e ma ke pa s ha lack assigned alues and o de ec anomalies and ou lie s by le e aging
machine lea ning algo i hms. In line wi h he indings o Indi a e al. (2023), ou app oach
emphasizes inco po a ing indus y-speci ic da a—speci ically he echnical ea u es o
pa s— o imp o e he accu acy o p ice p edic ions. This s udy aims o de elop a obus
me hodology ha accu a ely o ecas s op imal p ices and iden i ies i egula p icing pa e ns
ha could indica e e o s, poo da a quali y, o ex eme ma ke scena ios.
To achie e his, we compiled a comp ehensi e da ase ocusing on uel anks due o hei
c i ical ole in he ucking indus y and hei e enue po en ial. Gi en ha exis ing
me hodologies, as highligh ed by Spedica o e al. (2018) and Youbi e al. (2023), did no
p oduce sa is ac o y esul s o ou speci ic da ase , we adap ed hese echniques using ou
domain knowledge o de elop a sel -con ained me hodology ha is eadily applicable o eal-
wo ld and indus y-speci ic scena ios.
1.3 Resea ch con ibu ions and applica ion
This s udy con ibu es o he exis ing body o li e a u e by expanding on he me hodologies
p esen ed by Spedica o e al. (2018),Indi a e al. (2023), and Youbi e al. (2023), add essing
hei limi a ions by inco po a ing echnical ea u es ha a e c i ical o accu a e p icing in he
a e ma ke sec o . By ackling he esea ch ques ion o how machine lea ning algo i hms can
be u ilized o p edic he p ices o a e ma ke pa s wi h no assigned alues and de ec p icing
anomalies, his s udy p o ides a obus , da a-d i en amewo k ha can be eadily applied o
eal-wo ld scena ios in he a e ma ke indus y. By add essing he gaps and limi a ions
iden i ied in p e ious s udies, ou esea ch o e s p ac ical insigh s ha enhance p icing
s a egies, p o i abili y, and compe i i eness in he a e ma ke sec o .
2. Me hodology
The i s s ep was o ga he da a. The da ase was hen p ep ocessed, and key a ibu es we e
iden i ied. Selec ed a ibu es we e scaled, and di e en models we e compa ed o de e mine
which model was he bes . The Random Fo es Reg esso model was chosen, and i was hen
ained and es ed o p edic he p ices o he pa s. Figu e 1 depic s a block diag am o he en i e
Figu e 1. Block diag am o me hodology
Mode n Supply
Chain Resea ch
and Applica ions
413
me hodology. Web sc aping echniques, backend da a, and pa d awings we e used o c ea e
he da ase . I had hund eds o housands o da a poin s. Howe e ,a ew housand-pa numbe s
we e chosen om he la ge da ase o ocus on one speci ic uck assembly, he uel ank.
To ensu e da a quali y, a ibu es wi h insu icien da a poin s o lacking ele an
in o ma ion we e emo ed. Figu e 2 shows a ba cha ha was used o calcula e he pe cen age
o da a co e ed in each o he emaining a ibu es. A ibu es wi h less han 80% co e age in
he da a we e emo ed. Fu he mo e, all a ibu e alues we e cleaned o noise and
s anda dized o ollow he same o ma . Nume ical a ibu es wi h missing alues we e illed
wi h he median alue, whe eas ca ego ical a ibu es we e illed wi h he mos equen ly
occu ing ca ego ies. The numbe o a ibu es was educed by less han 60% as a esul o his.
Figu e 3 depic s he missing alue ma ix o hese a ibu es, whe e he whi e spaces be ween
he ma ix indica e he amoun o missing da a.
The equency o he p ice poin s was obse ed using a his og am in he ini ial analysis.
Based on his his og am, an a e age ange o p ices was de e mined, wi h p ices ou side o his
ange conside ed ou lie s. Figu e 4 depic s he accep able cos ange and he ou lie s.
To in es iga e he impac on p ices, ea u e enginee ing was ca ied ou by calcula ing new
a ibu es such as a ea, mass, and olume based on exis ing a ibu es such as diame e and
leng h. To imp o e he model’s pe o mance, new a ibu es such as shape and en we e
loaded and p ep ocessed in o he da ase . Figu e 5 shows a hea map c ea ed wi h he Seabo n
lib a y in Py hon o de e mine he signi ican a ibu es o p ice p edic ion.
Using he Random Fo es Reg esso model, a ious ea u e selec ion echniques we e
employed o e alua e he impac o a ibu es on p ice p edic ion. Va iance Th eshold,
Selec KBes , and Recu si e Fea u e Elimina ion wi h C oss-Valida ion (RFECV) we e he 3
echniques used o e alua e he a ibu es.
The Va iance Th eshold me hod, ep esen ed by Equa ion (i), is e ec i e a elimina ing
ea u es wi h low a iance, assuming hey ha e a minimal con ibu ion o he p edic i e
model. Du ing he selec ion p ocess, he me hod au oma ically iden i ies and emo es ze o
a iance ea u es.
selec o ¼Va ianceTh esholdðÞ (i)
Equa ion (ii) depic s he Selec KBes me hod, which uses a sco e unc ion called _ eg ession
and a pa ame e ko 13. The pa ame e k is se o 13, indica ing a p e e ence o keep he op 13
ea u es deemed mos in luen ial in p edic ing a e ma ke pa p ices. This me hod assesses
he s a is ical ela ionship be ween each ea u e and he a ge a iable, anking hem
acco ding o hei signi icance. The chosen pa ame e s s ike a balance be ween ea u e
ichness and model e iciency, aking in o accoun he linea ela ionship be ween ea u es and
he a ge a iable.
selec o ¼Selec KBes �sco e unc ¼ eg ession;k¼13�(ii)
The RFECV me hod, as shown in Equa ions (iii) and (i ), employs he Suppo Vec o
Reg esso (SVR) as an es ima o wi h a linea ke nel. This me hod sys ema ically e alua es
ea u e ele ance by ecu si ely emo ing he leas in o ma i e ea u es. The use o a linea
ke nel co esponds o he assumed linea ela ionship be ween ea u es and p ices in ou
da ase . S ep o 1 and c o 10 we e chosen as pa ame e s o ensu e ho ough ea u e e alua ion
while main aining compu a ional e iciency.
es ima o ¼SVRðke nel ¼linea Þ(iii)
selec o ¼RFECVðes ima o ;s ep ¼1;c ¼10Þ(i )
Based on ou da ase , RFECV elimina ion p oduced he mos a o able esul s, and Table 1
shows he ankings o all RFECV-calcula ed a ibu es. This in o ma ion was used o conduc
ial and e o es s on a ibu es o de e mine which con ibu ed o he highes accu acy.
MSCRA
6,4
414
Figu e 2. Pe cen age da a co e ed in each a ibu e
Mode n Supply
Chain Resea ch
and Applica ions
415
Recu si e Fea u e Elimina ion wi h C oss-Valida ion (RFECV) echnique ou pe o ms
Selec KBes and Va iance Th eshold because i can cap u e in ica e ela ionships and
dependencies be ween ea u es. By emo ing less ele an ea u es du ing c oss- alida ion,
RFECV excels a e alua ing he collec i e impac o ea u es in ou da ase , which includes
echnical a ibu es o au omobile pa s. The me hod e ains he mos ele an ea u es,
imp o ing accu acy and in e p e abili y.
Selec KBes , which is e icien a selec ing op ea u es based on indi idual me ics, does
no conside ea u e in e ac ions. Va iance Th eshold, which ocuses on a iance wi hin
indi idual ea u es, may miss impo an associa ions equi ed o accu a e p icing p edic ions.
Figu e 4. P ice dis ibu ion including accep able ange and ou lie s
Figu e 3. A ibu es included a e p ep ocessing
MSCRA
6,4
416
As a esul , RFECV’s conside a ion o ea u e in e ac ions and dependencies makes i mo e
sui able o selec ing impac ul ea u es and imp o ing p edic i e model accu acy.
Fea u e scaling was used o ensu e ha he ange o alues was uni o m. Nume ical alues
we e no malized, and ca ego ical da a was encoded using labels. No maliza ion was used o
adjus nume ical alues such as leng h, diame e , a ea, mass, and so on o a s anda d scale,
elimina ing po en ial biases caused by di e en measu emen uni s o scales. Ca ego ical da a,
on he o he hand, such as ma e ial, inish, shape, and so on, was encoded using labels,
Table 1. Ranking based on RFECV
Fea u es Ranking Fea u es Ranking
Leng h 2 An i-Siphon 16
Thickness 6 GROSS_WT 3
Ou e diame e 1 TOTAL_SALES_DEMAND 1
Inne diame e 1 A ea 1
Ma e ial 9 Volume 11
Finish 13 Mass 8
Moun ing loca ion 12 SHAPE 10
Fuel ank capaci y 4 INSTA_HEAT 14
Fille neck o end leng h 1 VENT 5
Ba les included 7 In eg al Fuel Tank 15
Sou ce(s): Au ho s’ own wo k
Figu e 5. Hea map co ela ing all he a ibu es
Mode n Supply
Chain Resea ch
and Applica ions
417
allowing o he ep esen a ion o quali a i e in o ma ion in a nume ical o ma sui able o
compu a ional analysis.
Gi en he la ge olume o a ailable da a, he da ase was di ided in o a 70% aining se and
a 30% es se . The alida ion da ase was assigned a nominal amoun (10%) o he aining
da ase . To es ima e he pe o mance o he models, he e o me ic Roo Mean Squa ed E o
(RMSE) was chosen.
Because o i s comp ehensibili y, abili y o cap u e p edic ion accu acy, and emphasis on
penalizing la ge e o s, he Roo Mean Squa e E o (RMSE) me ic s ands ou as a supe io
choice in a ious modeling scena ios. The a e age magni ude o he e o s be ween p edic ed
and ac ual alues is measu ed by RMSE, p o iding a s aigh o wa d unde s anding o how a
o he model’s p edic ions a e om he ue alues. Fu he mo e, because o i s squa ed
na u e, RMSE gi es signi ican weigh o la ge e o s, making he me ic mo e sensi i e o
ou lie s o ex eme de ia ions. Fu he mo e, RMSE is well-sui ed o eg ession- ype
p oblems in which he goal is o minimize p edic ion e o s.
A c i ical aspec o ou me hodology was he e alua ion o a ious eg ession models, such
as Random Fo es Reg ession, AdaBoos Reg ession, Bagging Reg ession, Suppo Vec o
Reg ession (SVR), and K-nea es Neighbo Reg ession. The accu acy on he alida ion se ,
mean RMSE sco e, and s anda d de ia ion we e all conside ed when selec ing a model.
No ably, he Random Fo es Reg esso eme ged as he supe io choice and his sec ion
in es iga es he limi a ions o al e na i e models while explaining why he Random Fo es
Reg esso was chosen.
2.1 Limi a ions o al e na i e eg ession models
O he algo i hms conside ed o his ask had limi a ions, bu Random Fo es eme ged as he
bes op ion. AdaBoos ’s sensi i i y o noisy da a and ou lie s could pose a p oblem o ou
da ase , which may ha e quali y a ia ions. Bagging, while e ec i e in educing a iance,
may impai in e p e abili y due o i s ensemble na u e. Fu he mo e, Suppo Vec o
Reg ession (SVR) necessi a es me iculous hype pa ame e uning, which inc eases he ime
equi ed o implemen a ion. Fo high-dimensional da ase s, K-Nea es Neighbo s (KNN) can
su e om he “cu se o dimensionali y,” which can ha e an impac on accu acy.
When dealing wi h la ge and high-dimensional da ase s, Random Fo es excels a
p edic i e analysis, especially when he e a e complex in e ac ions be ween ea u es o
nonlinea ela ionships wi h he a ge a iable. I pe o ms well e en wi hou ex ensi e
hype pa ame e uning and e ec i ely handles noisy da a. Random Fo es appea ed as he bes
op ion among he men ioned algo i hms because he goal is o ob ain obus p edic ions while
dealing wi h di e se ypes o da a and main aining good in e p e abili y, as shown in Table 2.
I is c i ical o accu a ely assess model pe o mance du ing he de elopmen p ocess.
Valida ion echniques help wi h his by assessing how well he Random Fo es Reg esso ( )
model gene alizes o new da a. In his s udy, he model was ained using he expec ed p ice
ange, and hen i s alida ion sco e was calcula ed. This s ep in ol ed de e mining how well
he model p edic ed p ices on da a ha had no been p e iously ained on. Howe e , mo e
Table 2. Compa ison o di e en models
Models Accu acy Mean RMSE sco e S anda d de ia ion
RandomFo es Reg esso 0.717507 61.800927 14.306699
AdaBoos Reg esso 0.451401 70.588170 11.112890
BaggingReg esso 0.673993 64.175979 14.116304
SVR 0.464940 113.950125 16.046542
KNeighbo sReg esso 0.622910 100.102034 16.436865
Sou ce(s): Au ho s’ own wo k
MSCRA
6,4
418
