Predicting stock returns with machine learning: Global versus sector models

Wi e , Johannes
A icle
P edic ing s ock e u ns wi h machine lea ning: Global
e sus sec o models
Junio Managemen Science (JUMS)
P o ided in Coope a ion wi h:
Junio Managemen Science e. V.
Sugges ed Ci a ion: Wi e , Johannes (2025) : P edic ing s ock e u ns wi h machine lea ning: Global
e sus sec o models, Junio Managemen Science (JUMS), ISSN 2942-1861, Junio Managemen
Science e. V., Planegg, Vol. 10, Iss. 3, pp. 561-581,
h ps://doi.o g/10.5282/jums/ 10i3pp561-581
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/326965
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/
Junio Managemen Science 10(3) (2025) 561-581
Junio Managemen Science
www.jums.academy
ISSN: 2942-1861
Edi o :
DOMINIK VAN AAKEN
Ad iso y Edi o ial Boa d:
FREDERIK AHLEMANN
JAN-PHILIPP AHRENS
THOMAS BAHLINGER
MARKUS BECKMANN
SULEIKA BORT
ROLF BRÜHL
KATRIN BURMEISTER-LAMP
CATHERINE CLEOPHAS
NILS CRASSELT
BENEDIKT DOWNAR
KERSTIN FEHRE
MATTHIAS FINK
DAVID FLORYSIAK
GUNTHER FRIEDL
MARTIN FRIESL
FRANZ FUERST
WOLFGANG GÜTTEL
NINA KATRIN HANSEN
ANNE KATARINA HEIDER
CHRISTIAN HOFMANN
SVEN HÖRNER
STEPHAN KAISER
NADINE KAMMERLANDER
ALFRED KIESER
ALEKSANDRA KLEIN
NATALIA KLIEWER
DODO ZU KNYPHAUSEN-AUFSESS
SABINE T. KÖSZEGI
ARJAN KOZICA
CHRISTIAN KOZIOL
MARTIN KREEB
WERNER KUNZ
HANS-ULRICH KÜPPER
MICHAEL MEYER
JÜRGEN MÜHLBACHER
GORDON MÜLLER-SEITZ
J. PETER MURMANN
ANDREAS OSTERMAIER
BURKHARD PEDELL
ARTHUR POSCH
MARCEL PROKOPCZUK
TANJA RABL
SASCHA RAITHEL
NICOLE RATZINGER-SAKEL
ASTRID REICHEL
KATJA ROST
THOMAS RUSSACK
FLORIAN SAHLING
MARKO SARSTEDT
ANDREAS G. SCHERER
STEFAN SCHMID
UTE SCHMIEL
CHRISTIAN SCHMITZ
MARTIN SCHNEIDER
MARKUS SCHOLZ
LARS SCHWEIZER
DAVID SEIDL
THORSTEN SELLHORN
STEFAN SEURING
VIOLETTA SPLITTER
ANDREAS SUCHANEK
TILL TALAULICAR
ANN TANK
ANJA TUSCHKE
MATTHIAS UHL
CHRISTINE VALLASTER
PATRICK VELTE
CHRISTIAN VÖGTLIN
BARBARA E. WEISSENBERGER
ISABELL M. WELPE
HANNES WINNER
THOMAS WRONA
THOMAS ZWICK
Volume 10, Issue 3, Sep embe 2025
JUNIOR
MANAGEMENT
SCIENCE
Johannes Wi e ,P edic ing S ock Re u ns Wi h Machine
Lea ning: Global Ve sus Sec o Models
Robin Roskosch, Bewa e o Bullshi –A Quali a i e S udy on
Young Adul s’ Sus ainabili y Awa eness o Online
Se ices
Nadhilla Mazaya,Boa d Gende Di e si y: E idence F om
Indonesia
Alexande Sake, Value C ea ion Oppo uni ies o Gene a i e
AI –A Case S udy
Jus us Olb ich, The E ec o Changes in In e nal Con ol
Sys ems on Audi Risk
Jan Oli e Ho s mann, Manda o y ESG Disclosu e and Fi m
Value –A Quan i a i e Analysis o he E ec o
Di ec i e 2014/95/EU on Fi m Value
Me e Anna Gläse , Go e nmen In e en ions Du ing he
COVID-19 Pandemic, Cul u e, and Co po a e Cos
Beha iou
Zewei Shi, Modeling he Impac o Emission C edi Sys ems on
Au omo i e P oduc Po olios: A Ma hema ical
Analysis o Policy E ec s in Eu ope, China, and he
U.S. Unde Di e en Demand Scena ios
Hagen Alexande Höne loh, Nume ical S udies o he
Scheduling o Con inuous Annealing Lines
Lea Wedel, KPIs o Sus ainabili y: De ining he S a egy o a
Sus ainable Fu u e in he Insu ance Indus y
561
582
609
631
657
677
715
748
781
810
Published by Junio Managemen Science e.V.
This is an Open Access a icle dis ibu ed unde he e ms o he CC-BY-4.0
(A ibu ion 4.0 In e na ional). Open Access unding p o ided by ZBW.
ISSN: 2942-1861
P edic ing S ock Re u ns Wi h Machine Lea ning: Global Ve sus Sec o Models
Johannes Wi e
Technical Uni e si y o Munich
Abs ac
Recen s udies highligh he supe io pe o mance o non-linea machine lea ning models, such as neu al ne wo ks, o e
adi ional linea models in p edic ing c oss-sec ional s ock e u ns. These models a e capable o cap u ing complex non-linea
in e ac ions be ween p edic i e signals and u u e e u ns. This hesis esea ches whe he sec o -speci ic neu al ne wo ks
can de ec sec o - ela ed ela ionships o ou pe o m a global neu al ne wo k. I e alua es he p edic i e powe o hese
models a he s ock le el and in po olios based on e u n o ecas s, cons uc ing long-sho po olios om he ne wo ks’
so ed p edic ions. A global neu al ne wo k model ained on he ull sample o s ocks domina es neu al ne wo ks ained
on indi idual GICS sec o s in p edic ing he c oss-sec ion o US s ock e u ns. Sec o -speci ic neu al ne wo ks ail o gain
an ad an age by cap u ing complex sec o -speci ic in e ac ions. They unde pe o m he global neu al ne wo k especially in
he ea ly ou -o -sample pe iod. The smalle sample size o each GICS sec o equi es a ade-o be ween model complexi y
and obus model es ima ion. Pooling he da a o he global model sol es his p oblem and suppo s he p edic i e powe o
neu al ne wo ks o s ock e u ns.
Keywo ds: c oss-sec ion o s ock e u ns; machine lea ning; neu al ne wo ks; e u n p edic ion; sec o models
1. In oduc ion
I compa e a global neu al ne wo k wi h sec o -speci ic
neu al ne wo ks o p edic he c oss-sec ion o US s ock e-
u ns. The e o e, I e alua e he p edic i e powe o he wo
di e en models a he s ock le el and in po olios con-
s uc ed based on e u n o ecas s. Recen esea ch demon-
s a es he abili y o non-linea machine lea ning models
such as neu al ne wo ks o ou pe o m adi ional linea
models in p edic ing he c oss-sec ion o e u ns. I show
how hese neu al ne wo ks pe o m be e when ained on
pooled da a ac oss sec o s han on sec o -speci ic da a om
GICS sec o s. My da a sample co e s he sample pe iod om
July 1963 o Decembe 2022, wi h an ou -o -sample pe iod
om Janua y 1994 o Decembe 2022. The global neu al
I would like o exp ess my g a i ude o all he people who ha e suppo ed
me in he ealiza ion o his hesis. In pa icula , I would like o hank my
supe iso , D . Ma hias Hanaue , o gi ing me he oppo uni y o wo k
on his in e es ing opic and o his guidance and inspi a ion h oughou
he wo k on my hesis.
ne wo k achie es a highe posi i e ou -o -sample R2
OOS han
he sec o models, which unde pe o m a nai e o ecas o
ze o. Fo long-sho po olios based on he so ed p edic-
ions o he neu al ne wo ks, he global model ou pe o ms
he sec o models in e ms o gene a ed mon hly e u ns
and Sha pe a io. The compa ison o sec o neu al ne wo ks
wi h simple OLS models highligh s he necessa y ade-o
be ween es ima ing a s able model wi hou o e i ing and
cap u ing complex sec o -speci ic in e ac ions.
I ollow he machine lea ning aining app oach o Gu e
al. (2020). I ain neu al ne wo ks wi h h ee hidden lay-
e s on 212 s ock-speci ic signals om p io li e a u e o p e-
dic he c oss-sec ion o e u ns o he US s ock ma ke . The
neu al ne wo ks a e ained using a ecu si e scheme wi h
inc easing aining samples and a ixed size olling sample
o alida ion. I e i all models once pe yea in Decembe
and p edic mon hly ou -o -sample e u ns o e he ollowing
yea .
In o al, I ain ele en di e en neu al ne wo k models.
A global model is ained on he ull sample o s ocks. The
DOI: h ps://doi.o g/10.5282/jums/ 10i3pp561-581
© The Au ho (s) 2025. Published by Junio Managemen Science.
This is an Open Access a icle dis ibu ed unde he e ms o he CC-BY-4.0
(A ibu ion 4.0 In e na ional). Open Access unding p o ided by ZBW.
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581562
sec o model consis s o en sec o -speci ic neu al ne wo ks
ained on subse s o he sample da a il e ed o each o he
en sec o s de ined by he Global Indus y Classi ica ion S an-
da d (GICS). To assign s ocks o sec o s, I use exis ing da a
on GICS classi ica ions and a cus om mapping om he S an-
da d Indus ial Classi ica ion (SIC) sys em o GICS sec o s.
Fi s , I e alua e he wo models on hei he ou -o -sample
p edic i e pe o mance o indi idual s ock e u n o ecas s.
The global model ou pe o ms he sec o models in all indi-
idual GICS sec o s and in he ull sample. The global neu al
ne wo k p oduces a mon hly R2
OOS o 3.37% in he ull sam-
ple and a posi i e R2
OOS o each o he en sec o s. The sec o
models achie e a mon hly R2
OOS o -6.06% in he ull sample,
so hey unde pe o m a nai e o ecas o ze o o all mon hly
s ock e u ns. Sec o models pe o m pa icula ly poo ly in
p edic ing s ock e u ns in sec o s wi h small sample sizes.
To unde s and he di e ences in p edic i e pe o mance,
I examine he impo ance o each inpu a iable in p edic ing
e u ns wi h he neu al ne wo k models. Va iable impo ance
is de e mined by he educ ion in R2
OOS ha esul s om se -
ing all alues o a pa icula signal o ze o while holding
all o he model es ima es ixed. The di e en models sha e
mos o hei mos impo an a iables. Analogous o Bli z
e al. (2023), Size is he mos in luen ial signal in p edic -
ing e u ns in he global model and in mos sec o models.
The neu al ne wo ks end o pe o m well in ou -o -sample
e u n o ecas ing when hei ela i e impo ance is skewed
owa ds Size. Sec o models wi h a b oade se o in luen ial
cha ac e is ics unde pe o m ou -o -sample.
I compa e he p o i abili y o po olios based on he
so ed p edic ions o he global model wi h he sec o mod-
els. A he end o each mon h, I so he s ocks in o decile
po olios and calcula e he alue-weigh ed e u ns o hold-
ing he decile po olios o e he nex mon h. A long-sho
po olio buys s ocks wi h he highes expec ed e u ns and
sells hose wi h he lowes . Po olio s a egies based on he
ou -o -sample p edic ions o he global model ou pe o m
he sec o -speci ic models. A long-sho po olio based on
he global model’s o ecas s achie es an a e age mon hly
ou -o -sample e u n o 2.71% and an annualized Sha pe
a io o 2.07. A long-sho po olio based on he o ecas s
o he sec o models gene a es an ou -o -sample mon hly e-
u n o 0.99% and an annualized Sha pe a io o 1.14. The
ou pe o mance o he global neu al ne wo k is pa icula ly
s ong in he ea ly yea s o he ou -o -sample pe iod. Du ing
his pe iod, he global model achie es i s highes e u ns
while he sec o models s uggle o emain p o i able.
The sec o -speci ic neu al ne wo ks con inue o unde pe -
o m when compa ed o simple o dina y leas squa es (OLS)
models. A long-sho po olio based on he so ed p edic-
ions o he OLS sec o models gene a es a highe alue-
weigh ed e u n han he sec o neu al ne wo ks. Howe e ,
he OLS ou pe o mance comes only om he i s hal o
he ou -o -sample pe iod when less aining da a is a ailable.
Small sec o sample sizes equi e a ade-o be ween s able
model es ima ion and cap u ing complex sec o -speci ic in-
e ac ions. The global neu al ne wo k ained on pooled da a
signi ican ly ou pe o ms he OLS models.
The global neu al ne wo k shows some ou -o -sample sec-
o alloca ion powe . In he c oss-sec ion o sec o s, i co -
ec ly p edic s highe ela i e e u ns o he mos p o i able
sec o s and lowe e u ns o he leas p o i able sec o s. As
a esul , he e u ns gene a ed by he global model a e lowe
wi h sec o -neu al po olios.
Sec ion 2 e iews he ecen li e a u e on machine lea n-
ing o e u n o ecas ing and global e sus indus y-speci ic
models. Sec ion 3p esen s he sou ces o s ock da a and in-
pu signals o e u n p edic ion and explains he sec o clas-
si ica ions. Sec ion 4desc ibes he me hodology used o ain
neu al ne wo ks o e u n p edic ion and o cons uc po -
olios based on hese p edic ions. Sec ion 5p esen s he e-
sul s o compa ing he o ecas ing pe o mance o he global
model wi h he sec o models. Sec ion 6concludes.
2. Li e a u e e iew
The pas decades p oduced a a ie y o li e a u e ocus-
ing on p edic ing he c oss-sec ion o s ock e u ns. Au ho s
explo e a a ie y o a iables in linea models, bu he e is
s ill a lack o consensus ega ding which a iables a e e-
la ed o expec ed s ock e u ns. This p oblem is o en e-
e ed o as ac o zoo. Linea models canno deal wi h many
a iables and hei po en ial nonlinea i ies and in e ac ions.
The e o e, ecen esea ch ocuses on mo e complex machine
lea ning models o handle he high dimensionali y in he ac-
o zoo.
Ea ly li e a u e ocuses on single machine lea ning mod-
els and hei abili y o ou pe o m adi ional me hods.
Mo i z and Zimme mann (2016) p opose ee-based con-
di ional po olio so s as a machine lea ning app oach. In
hei models, ecen pas e u ns wi hin he las six mon hs
p edic u u e e u ns and ou pe o m linea models like
Fama-MacBe h eg essions. Excess e u ns pe sis e en a e
accoun ing o ansac ion cos s and common isk ac o s.
T adi ional me hodologies wi h linea assump ions ail o
cap u e a nonlinea ela ionship be ween pas and u u e
e u ns.
A nonpa ame ic model using adap i e g oup leas abso-
lu e sh inkage and selec ion ope a ion (LASSO) isola es el-
e an p edic o s in a high-dimensional se ing in F eybe ge
e al. (2020). A small subse o a iables, including size, o-
al ola ili y, and ecen e u n-based me ics, p o ide unique
p edic i e powe . Thei model signi ican ly ou pe o ms lin-
ea app oaches like hose o Lewellen (2015) wi h highe
ou -o -sample Sha pe a ios. The nonpa ame ic model se-
lec s ewe a iables in-sample han he linea models bu
cap u es nonlinea in e ac ions.
Gu e al. (2020) a e among he i s o p esen a com-
pa a i e analysis o machine lea ning models o p edic ing
s ock e u ns. Thei models ag ee on a small se o domi-
nan a iables, wi h p ice ends, liquidi y, and ola ili y as
he mos in luen ial p edic o s. Neu al ne wo ks pe o m
he bes among all machine lea ning models, and po o-
lios so ed on neu al ne wo k e u n p edic ions double he
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581 563
Sha pe a ios o linea models. Shallow neu al ne wo ks ou -
pe o m deepe ones due o he limi ed da a and low signal-
o-noise a io in empi ical asse p icing. In e ac ions and
nonlinea ela ions be ween a iables d i e he ou pe o -
mance o machine lea ning me hods. Aze edo and Hoegne
(2023) epo simila esul s wi h ee-based and neu al ne -
wo k app oaches. Thei machine lea ning me hods unco e
in e ac ion e ec s challenging adi ional isk-based expla-
na ions in asse p icing. Neu al ne wo k models u ilize 299
s ock ma ke anomalies o achie e mon hly ou -o -sample e-
u ns 1% highe han a linea benchma k. Linea eg essions
a e easy o in e p e bu unde pe o m hei machine lea n-
ing models using s a is ical signi icances and e u ns.
Mos ecen li e a u e uses echnically mo e ad anced
machine lea ning s a egies. Aze edo e al. (2024) examine
he expec ed e u ns o deep lea ning s a egies. Thei long
sho - e m memo y (LSTM) models yield ne e u ns o up o
1.42% pe mon h, e en a e accoun ing o he ecen e a o
high liquidi y, ansac ion cos s, and pos -publica ion decay.
S a egies combining se e al machine lea ning models con-
s an ly achie e signi ican e u ns a e cos . Cos mi iga ion
echniques educe u no e and ading cos s bu do no im-
p o e ne pe o mance. L. Chen e al. (2024) combine h ee
deep neu al ne wo k models wi h no-a bi age cons ain s
o es ima e asse p icing models o US s ocks. Inco po a -
ing speci ic domain knowledge in o he echnical implemen-
a ion enhances p edic ion accu acy and ou -o -sample pe -
o mance. Thei deep lea ning s a egies wi h no-a bi age
cons ain s ou pe o m o he machine lea ning benchma ks
in Sha pe a io and iden i y he co e a iables d i ing asse
p ices.
O he esea ch ocuses on using machine lea ning o p e-
dic s ock e u ns no only in he US bu globally. Tobek
and H onec (2021) agg ega e 153 anomalies ac oss global
ma ke s in o one misp icing signal using machine lea ning.
Thei s a egy ou pe o ms linea models ou -o -sample in
a ious in e na ional ma ke s. Ex ending he aining sam-
ple wi h in e na ional da a does no imp o e ou -o -sample
pe o mance o he US ma ke . Howe e , machine lea ning
models ained on US s ocks pe o m well in ma ke s ou side
he US. Cakici e al. (2023) in es iga e machine lea ning’s
c oss-sec ional e u n p edic abili y ac oss 46 global s ock
ma ke s. Combining p edic ions om mul iple machine
lea ning models deli e s obus ou -o -sample e u ns ac oss
di e se ma ke s. De eloped ma ke s show highe p o i abil-
i y han eme ging ones. Fi m size and idiosync a ic isk a e
he mos impo an a iables o p edic ions, wi h highe
e u ns in smalle i ms and ma ke s wi h mo e idiosync a ic
isk. Aze edo e al. (2023) also ocus on he ou -o -sample
pe o mance o di e en machine lea ning models ac oss an
in e na ional da a sample. Neu al ne wo ks and composi e
p edic o s pe o m he bes . These models achie e signi -
ican mon hly long-sho e u ns o a ound 2%. Po olio
e u ns emain signi ican e en a e ansac ion cos s and
ou pe o m linea benchma k models. D obe z and O o
(2021) use machine lea ning s a egies o p edic Eu opean
s ock e u ns. Like in he US ma ke , machine lea ning
models ou pe o m adi ional linea models by cap u ing
nonlinea i ies and a iable in e ac ions. Neu al ne wo ks
and classi ica ion-based app oaches pe o m bes and gene -
a e signi ican e u ns e en a e ansac ion cos s. Suppo
ec o machines, which classi y s ocks in o decile po olios,
deli e e en highe e u ns by elimina ing he noise o ex-
pec ed e u ns a he s ock le el. Leippold e al. (2022) apply
machine lea ning models o he Chinese s ock ma ke . In a
ma ke domina ed by e ail in es o s, liquidi y and ola ili y
indica o s ha e p edic i e powe o e adi ional a iables
like alua ion a ios. Neu al ne wo ks pe o m bes , pa icu-
la ly o small-cap and non-s a e-owned i ms. They achie e
highe p edic abili y in China han in he US due o dis inc
asse p icing dynamics d i en by local in es o beha io .
Hanaue and Kalsbach (2023) assess a ious machine lea n-
ing models o p edic ing s ock e u ns in a b oad sample o
eme ging ma ke s. Like in de eloped ma ke s, hei mod-
els iden i y nonlinea i ies and in e ac ions among a iables.
T ee-based me hods and neu al ne wo ks deli e supe io
long-sho e u ns and alphas o e linea models. E icien
ading ules ensu e machine lea ning p edic ions ou pe -
o m e en a e ansac ion cos s, sho -selling cons ain s,
and limi ing he sample o big s ocks.
Despi e he s ong pe o mance o machine lea ning mod-
els o s ock ma ke p edic ion, he e a e s ill p oblems in
implemen ing hem in p ac ice. Rasekhscha e and Jones
(2019) ocus on mi iga ing o e i ing in machine lea ning
models. Fea u e enginee ing and o ecas combina ions de-
c ease he isk o o e i ing. These echniques inc ease he
signal- o-noise a io and p oduce mo e obus p edic ions.
A amo e al. (2023) c i icize high limi s- o-a bi age en-
i onmen s and exclude s ocks like mic ocaps and dis essed
i ms. Machine lea ning po olios o en ely on long and
sho posi ions ha a e impossible in p ac ice. The p o i abil-
i y o machine lea ning s a egies is educed when ading
cos s a e conside ed due o high u no e . The au ho s p o-
pose including ading cos s in machine lea ning models and
imposing economic es ic ions. Bli z e al. (2023) epo he
impac o a ying p edic ion ho izons in machine lea ning
models o s ock e u n p edic ions. While one-mon h o e-
cas s yield high g oss e u ns, he ne e u ns conside ing
ansac ion cos s a e minimal a e 2004. Machine lea ning
models wi h longe p edic ion ho izons p o ide signi ican
ne alpha due o educed po olio u no e . One-mon h o e-
cas s ely on sho - e m p ice signals, whe eas longe ho i-
zon p edic ions ely mo e on alue-o ien ed signals. Aligning
he design o machine lea ning models wi h ading ho izons
enhances p o i abili y h ough educed u no e and be e
a e -cos pe o mance.
In his hesis, I ocus on sec o -speci ic e sus global
machine lea ning models. The e o e, i is app op ia e o
conside p io esea ch on he ela ion be ween indus y-
speci ic and ma ke p edic ions. This includes adi ional
me hods o p edic ing he c oss-sec ion o s ock e u ns, like
ac o in es ing and o he linea eg ession models. Kim e
al. (2013) enhance he linea esidual income model (RIM)
wi h indus y-speci ic ac o s using he alue- o-book (V/B)
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581564
a io. Decomposing V/B in o indus y-speci ic componen s
be e p edic s u u e abno mal e u ns and ou pe o ms a-
di ional RIM implemen a ions. Thei indus y-adjus ed RIM
model p o ides supe io p edic i e accu acy o abno mal
e u ns compa ed o con en ional alua ion measu es. Liu
e al. (2014) analyze he p edic i e powe o indus y e -
ec s in op ion-implied ola ili y measu es on s ock e u ns.
Indus y-neu al po olios ou pe o m ull-uni e se po -
olios wi h highe Sha pe a ios and lowe downside isk.
Ca aglia e al. (2006) explo e egion-neu al and indus y-
neu al alue po olios. Indus y-neu al po olios o e
mo e s able e u ns and highe in o ma ion a ios due o
lowe ola ili y and less cyclicali y. They cap u e he global
alue p emium mo e e ec i ely and p o ide a be e isk-
e u n p o ile.
Mode n machine lea ning models also cap u e econom-
ically meaning ul in e dependencies among indus ies. Ra-
pach e al. (2019) examine c oss-indus y e u n p edic abil-
i y wi h machine lea ning models like LASSO. Due o g adual
in o ma ion di usion, lagged e u ns om indus ies like i-
nancial and commodi y sec o s can p edic e u ns in o he
sec o s. An indus y- o a ion s a egy based on hese c oss-
indus y signals ou pe o ms linea me hods. Cine (2019)
uses a andom o es model as a machine lea ning s a egy
o p edic ma ke e u ns wi h indus y e u ns. Indus y e-
u ns p o ide signi ican ou -o -sample p edic i e powe o
he ma ke index. Random o es s ou pe o m adi ional
linea models due o hei capaci y o cap u e bo h linea
and non-linea dynamics. Indus y-le el in o ma ion o e-
cas s ma ke mo emen s in a way ha linea models ail o
cap u e.
This hesis ex ends he exis ing li e a u e on machine
lea ning in empi ical asse p icing by in eg a ing i wi h
indus y-speci ic s a egies. Building upon he wo k o Gu
e al. (2020), I use neu al ne wo ks as well-pe o ming ma-
chine lea ning models o o ecas he c oss-sec ion o s ock
e u ns o he US ma ke . I compa e long-sho po olio
e u ns achie ed by sec o -speci ic models wi h hose o a
global model.
3. Da a
3.1. S ock da a
My sample includes all US s ocks lis ed on he NYSE,
AMEX, and NASDAQ. The sample pe iod uns om July
1963 o Decembe 2022. I sou ce equi y e u ns and o he
s ock ma ke da a om CRSP. Accoun ing da a o eplica e
he Fama and F ench (2015) i e- ac o model is om Com-
pus a .
I calcula e mon hly excess e u ns as he one-mon h s ock
e u n om CRSP o e he isk- ee a e p o ided on Kenne h
R. F ench’s homepage.1To p edic he c oss-sec ion o s ock
e u ns, I sub ac he mon hly c oss-sec ional median excess
1See h p://mba. uck.da mou h.edu/pages/ acul y/ken. ench/da a_lib
a y.h ml (2024).
e u n om he mon hly excess e u n o each s ock. “Re-
u ns” h oughou his hesis deno e hese ela i e mon hly
s ock e u ns wi h he ma ke componen al eady emo ed.
The inpu a iables o he machine lea ning models come
om A. Y. Chen and Zimme mann (2022). I download he
Augus 2023 e sion o signals om hei Open Sou ce Asse
P icing (OSAP) websi e.2This includes 209 i m-le el cha -
ac e is ics eplica ed om he academic asse p icing li e a-
u e. I use he e ms a iables, signals, and cha ac e is ics
in e changeably h oughou his hesis.
In addi ion o he 209 signals om he OSAP da a, I de ine
h ee inpu a iables om CRSP da a. Sho - e m e e sal is
he p io one-mon h e u n, P ice is he na u al loga i hm o
he CRSP p ice da a ield, and Size is he na u al loga i hm
o he p ice mul iplied by he sha es ou s anding.
All inpu a iables a e signed so ha highe alues co e-
spond o highe expec ed e u ns. Following Gu e al. (2020),
Bli z e al. (2023) and o he ecen li e a u e, I ank all inpu
a iables c oss-sec ionally o each mon h in o he in e al o
[-1,1]. This helps neu al ne wo k models deal wi h a ying
anges o alues and di e en a iances ac oss he signals
du ing aining. Missing alues a e illed wi h he mon hly
c oss-sec ional median ank.
I ollow Hou e al. (2020) and Bli z e al. (2023) and ex-
clude mic ocaps om my sample o p e en hem om d i -
ing my esul s. I de ine mic ocaps as all s ocks wi h a mon hly
ma ke capi aliza ion below he 20 h pe cen ile o he NYSE
ma ke capi aliza ion in ha mon h.
A e excluding mic ocaps, my ull sample om July
1963 o Decembe 2022 includes app oxima ely 1.3 million
mon hly s ock obse a ions wi h a mon hly a e age o 1842
s ocks.
3.2. Sec o da a
To ain sec o -speci ic machine lea ning models, I assign
all s ocks in my sample o sec o s acco ding o he Global
Indus y Classi ica ion S anda d (GICS) om MSCI and S&P.
I ca ego ize s ocks in o en sec o s de ined by GICS: Ene gy,
Ma e ials, Indus ials, Consume Disc e iona y, Consume
S aples, Heal h Ca e, Financials, In o ma ion Technology,
Communica ion Se ices, and U ili ies. Due o he small
numbe o obse a ions, I include he Real Es a e sec o in
he Financials sec o . This co esponds o he GICS classi i-
ca ion be o e 2016, be e ep esen ing he bigges pa o
my sample pe iod.
I p e e he GICS o e he S anda d Indus ial Classi ica-
ion (SIC) sys em o en used in o he li e a u e on indus y
speci ics. The GICS is he mo e mode n indus y axonomy
wi h a s onge ocus on new indus ies like he compu e ,
so wa e, and in o ma ion echnology sec o s.
I sou ce da a on SIC indus y classi ica ion om CRSP
and da a on GICS sec o classi ica ion om Compus a . GICS
da a has weake co e age han SIC da a, pa icula ly om
2See h ps://www.openasse p icing.com/augus -2023-da a- elease/
(2023).

J. Wi e /Junio Managemen Science 10(3) (2025) 561-581 565
he beginning o my sample un il 1985. The e o e, I de elop
a mapping om SIC indus ies o GICS sec o s desc ibed in
Table 1. I a mon hly s ock obse a ion is no assigned o a
sec o unde GICS bu o a SIC indus y, I ill in he missing
GICS classi ica ion based on he mapping in Table 1.
The mapping is based on wo di e en inpu s. Fi s , I con-
side o e laps be ween he SIC and GICS classi ica ions o
s ock obse a ions wi h bo h da a poin s in my sample. Fo
example, 90% o all s ocks classi ied as Financials in SIC a e
also classi ied as Financials in GICS. Simila ly, o e 80% o all
s ocks classi ied as u ili ies in SIC a e also classi ied as U ili-
ies in GICS. Second, I conside he de ini ions o he indus y
axonomy in bo h classi ica ion sys ems.3Fo example, SIC
includes anspo a ion, communica ions, elec ic, gas, and
sani a y se ices in one Di ision. GICS classi ies T anspo a-
ion as an indus y g oup in he sec o Indus ials, commu-
nica ions belong o he sec o Communica ion Se ices, and
elec ic, gas, and sani a y se ices belong o he sec o U il-
i ies. This enables me o map he co esponding SIC majo
g oups o he GICS sec o s.
Table 1includes he en GICS sec o s used o sec o -
speci ic machine lea ning models. I summa izes all SIC digi
codes and he co esponding indus y desc ip ion mapped o
each GICS sec o . Fo example, I map he SIC codes 1200-
1399 (Majo G oup Coal Mining and Majo G oup Oil And
Gas Ex ac ion) and 2900-2999 (Majo G oup Pe oleum Re-
ining And Rela ed Indus ies) o he GICS sec o Ene gy. The
SIC indus ies mapped o a GICS sec o can be ei he en i e
Di isions (de e mined by a capi al le e ) o mo e de ailed
Majo G oups (de e mined by he i s wo digi s) and Indus-
y G oups (de e mined by he i s h ee digi s).
I I ecognize no clea ela ionship be ween a SIC indus y
and GICS sec o s, hen I don’ include his SIC indus y in he
mapping. A s ock wi h missing GICS da a and a SIC code no
con ained in he mapping will no be assigned o a GICS sec-
o . The same happens o s ocks wi h missing da a o bo h
classi ica ion sys ems. A e applying he mapping, 15,239
ou o 1.3 million mon hly s ock obse a ions in he ull sam-
ple a e no assigned o a GICS sec o . They a e included in
he aining da a o he global neu al ne wo k bu no in he
aining da a o he sec o -speci ic neu al ne wo ks.
4. Me hodology
4.1. Re u n p edic ion using machine lea ning
My me hodology ollows Gu e al. (2020) and Hanaue
and Kalsbach (2023), wi h he di e ence ha I ain one
global model and one sec o -speci ic model o each o he
en GICS sec o s.
I aim o p edic he c oss-sec ion o US s ock e u ns, so
I o ecas he ou pe o mance o a s ock ela i e o he US
s ock ma ke . The ela i e e u n o a s ock is de ined as
3See h ps://www.msci.com/ou -solu ions/indexes/gics o GICS de ini-
ions and h ps://www.osha.go /da a/sic-manual o SIC de ini ions.
el
i, = i, −Mk , (1)
whe e i, is he excess e u n o s ock iin mon h and Mk
is he c oss-sec ional median excess e u n ac oss all s ocks
in he sample in mon h .
I desc ibe he one-mon h-ahead ela i e e u n o a s ock
el
i, +1as an addi i e p edic ion e o model:
el
i, +1=E  el
i, +1|xi, +εi, +1. (2)
E [ el
i, +1|xi, ]is he condi ional expec ed ela i e e u n
o s ock iin mon h o mon h +1. I is condi ional as
i depends on xi, ∈Rp, a ec o o s ock-speci ic pinpu
a iables known a mon h .εi, +1is he p edic ion e o
e m.
I es ima e he expec ed ela i e e u n wi h he unknown
unc ion ∗, ∗:Rp→R. I es ima es he expec ed e u n
depending only on he ec o o ps ock-speci ic inpu a i-
ables a ailable in mon h :
E  el
i, +1|xi, = ∗(xi, ). (3)
In he case o neu al ne wo ks, he unknown unc ion
∗(x)is app oxima ed by a nonlinea unc ion (x,θ,ρ).
This unc ion is pa ame ized by a ec o o coe icien s θ
and a se o hype pa ame e s ρ. When aining neu al ne -
wo ks, he coe icien s θa e es ima ed om he aining da a
wi h espec o he hype pa ame e s ρand a p ede ined loss
unc ion L. The hype pa ame e s ρa e op imized conce n-
ing he loss unc ion Lbased on he es ima ed coe icien s θ
and a ailable da a.
Neu al ne wo ks as a o m o supe ised machine lea n-
ing ou pe o m linea models in p io li e a u e (Aze edo &
Hoegne , 2023; Gu e al., 2020). The e o e, I choose h ee-
laye neu al ne wo ks as he machine lea ning model o his
hesis. Appendix A4 desc ibes he model a chi ec u e and
hype pa ame e s used o ain he neu al ne wo ks. In ad-
di ion, I la e use o dina y leas squa es (OLS) models as a
benchma k o sec o -speci ic neu al ne wo ks.
The global neu al ne wo k model akes he ull sample o
1.3 million mon hly s ock obse a ions as inpu . The sec o -
speci ic machine lea ning models o he en GICS sec o s
ake all mon hly s ock obse a ions assigned o he espec-
i e GICS sec o as inpu . The e o e, he samples used o
ain he sec o -speci ic neu al ne wo ks di e signi ican ly
in hei numbe o obse a ions, depending on he sec o ’s
size.
To a oid da a leakage, I di ide all inpu samples in o h ee
disjoin ime pe iods, which always keep he empo al o de -
ing o he da a: he aining, alida ion, and es ing samples.
Fi s , I es ima e he neu al ne wo k coe icien s o a ange
o hype pa ame e alues on he aining sample. The ali-
da ion sample compa es he loss unc ion esul s o each se
o hype pa ame e s based on he es ima ed model om he
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581566
Table 1: Mapping o SIC o GICS indus ies
This able maps indus ies classi ied unde he S anda d Indus ial Classi ica ion (SIC) sys em o sec o s classi ied unde he Global Indus y Classi ica ion
S anda d (GICS) sys em used in his hesis. The i s column con ains he 11 di e en sec o s o he GICS sys em. Th oughou his hesis, he 11 h sec o ,
Real Es a e, is no conside ed sepa a ely, bu is included in he Financials sec o . The second and hi d columns con ain he SIC indus ies mapped o he
GICS sec o in he i s column. The second column con ains he SIC digi code, and he hi d includes he co esponding indus y desc ip ion. The SIC
indus ies mapped o a GICS sec o can be ei he en i e Di isions (de e mined by a capi al le e ) o mo e de ailed Majo G oups (de e mined by he i s
wo digi s) and Indus y G oups (de e mined by he i s h ee digi s). This mapping is used o classi y indi idual s ocks in o sec o s when GICS sec o
in o ma ion is una ailable in Compus a . I a SIC indus y classi ica ion is a ailable, he GICS sec o is added acco ding o his mapping. I no SIC
classi ica ion is a ailable ei he , he GICS sec o alue is ’Missing’, and he s ock is no included in he aining da a o he sec o models.
GICS Sec o SIC Code SIC Desc ip ion
10 - Ene gy 1200–1399 Coal Mining and Oil/Gas Ex ac ion
2900–2999 Pe oleum Re ining and Rela ed Indus ies
15 - Ma e ials Di ision B 1000–1499
(excluding 1200–1399)
Mining (excluding Coal Mining and Oil/Gas Ex ac ion)
2400–2499 Lumbe and Wood P oduc s, Excep Fu ni u e
2600–2699 Pape and Allied P oduc s
3300–3399 P ima y Me al Indus ies
20 - Indus ials Di ision C 1500–1799 Cons uc ion
Di ision E 4000–4999
(excluding 4800–4999)
T anspo a ion (excluding Communica ions and U ili ies)
Di ision J 9100–9999
(excluding 9900–9999)
Public Adminis a ion (excluding Nonclassi iable Es ablishmen s)
3400–3499 Fab ica ed Me al P oduc s, Excep Machine y and T anspo a ion
Equipmen
3500–3599 Indus ial and Comme cial Machine y
7320–7329 C edi Repo ing and Collec ion
7340–7349 Se ices o Dwellings and O he Buildings
7360–7369 Pe sonnel Supply Se ices
7390–7399 Miscellaneous Business Se ices
7500–7599 Au omo i e Repai Se ices and Pa king
7600–7699 Miscellaneous Repai Se ices
8710–8719 Enginee ing A chi ec u al and Su eying Se ices
8740–8749 Managemen and Public Rela ions
8900–8999 Se ices No Elsewhe e Classi ied
25 - Consume
Disc e iona y
Di ision G 5200–5999
(excluding 5400–5499)
Re ail T ade (excluding Food S o es)
Di ision F 5000–5199
(excluding 5140–5189,
5180–5189)
Wholesale T ade (excluding G oce ies and Bee , Wine, and Dis illed
Alcoholic Be e ages)
1500–1599 Building Cons uc ion Gene al Con ac o s and Ope a i e Builde s
2200–2299 Tex ile Mill P oduc s
2300–2399 Appa el and O he Finished P oduc s Made om Fab ics and Simila
Ma e ials
2500–2599 Fu ni u e and Fix u es
3100–3199 Lea he and Lea he P oduc s
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581 567
Table 1— con inued
GICS Sec o SIC Code SIC Desc ip ion
25 - Consume
Disc e iona y
3900–3999 Miscellaneous Manu ac u ing Indus ies
7000–7099 Ho els, Rooming Houses, Camps, and O he Lodging Places
7200–7299 Pe sonal Se ices
7800–7899 Mo ion Pic u es
7900–7999 Amusemen and Rec ea ion Se ices
30 - Consume S aples Di ision A 0100–0999 Ag icul u e, Fo es y, and Fishing
2000–2199 Food P oduc s and Tobacco P oduc s
5140–5159 Wholesale T ade - G oce ies
5180–5189 Wholesale T ade - Bee , Wine, and Dis illed Alcoholic Be e ages
5400–5499 Food S o es
35 - Heal h Ca e 2800–2899 Chemicals and Allied P oduc s (including D ugs)
3840–3849 Su gical Medical and Den al Ins umen s and Supplies
3850–3859 Oph halmic Goods
8000–8099 Heal h Se ices
8300–8399 Social Se ices
8730–8739 Resea ch De elopmen and Tes ing Se ices
9900–9999 Nonclassi iable Es ablishmen s
40 - Financials Di ision H 6000–6799 Finance, Insu ance, and Real Es a e
45 - In o ma ion
Technology
3570–3579 Compu e and O ice Equipmen
3600–3699 Elec onic and O he Elec ical Equipmen and Componen s
3820–3829 Labo a o y Appa a us and Analy ical Op ical Measu ing and Con olling
Ins umen s
7370–7379 Compu e P og amming Da a P ocessing
50 - Communica ion
Se ices
4800–4899 Communica ions
55 - U ili ies 4900–4999 Elec ic, Gas, and Sani a y Se ices
60 - Real Es a e Included in GICS sec o 40 - Financials
aining sample. The op imal hype pa ame e se minimizes
he loss unc ion on he alida ion sample and is hen used
o e ain i e di e en neu al ne wo ks on he aining sam-
ple. I use hese i e models o p edic he mon hly e u ns
o he es sample. The inal p edic ion o each s ock is he
a e age o e he i e indi idual model p edic ions o educe
he a iance in single o ecas s.
Following he aining app oach as in Gu e al. (2020)
and Bli z e al. (2023), I e ain he models once a he end
o e e y yea bu p edic e e y mon h using he la es model
and da a. The i s 18 yea s o my sample (July 1963 o De-
cembe 1981) a e he i s aining sample, and he nex 12
yea s (Janua y 1982 o Decembe 1993) he i s alida ion
sample. The i s one-yea es sample is he ollowing 12
mon hs, so he i s ou -o -sample (OOS) p edic ion is made
o Janua y 1994. To p edic he mon hly e u ns om Jan-
ua y 1995 o Decembe 1995, I ex end he aining sample by
one yea (July 1963 o Decembe 1982) and oll o wa d he
alida ion sample by one yea (Janua y 1983 o Decembe
1994). I epea his p ocedu e o each yea in my sample.
No u u e in o ma ion is leaked om a p e ious pe iod.
To e alua e he p edic i e pe o mance o indi idual
s ock e u n o ecas s on he es sample, I use he pooled
ou -o -sample R2
OOS de ined by Gu e al. (2020):
R2
OOS =1−PT
PN
i el
i, −ˆ
el
i, 2
PT
PN
i el
i, 2. (4)
This me ic compa es he ou -o -sample o ecas s wi h a
nai e o ecas o ze o, be e sui ed o indi idual e u n p e-
dic ions han he ypical o ecas wi h mean e u ns.
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581568
4.2. Va iable impo ance
As a p ima y measu e o in e p e he esul s o he ma-
chine lea ning models, I ank he espec i e inpu a iables
acco ding o hei a iable impo ance. This aims o iden i y
cha ac e is ics ha in luence he c oss-sec ion o expec ed e-
u ns. Following Gu e al. (2020) and Bli z e al. (2023), I
de ine a iable impo ance as he educ ion in panel p edic-
i e ou -o -sample R2
OOS. Fo each annually ained neu al
ne wo k, I i e a i ely se he alues o each inpu a iable o
ze o while holding he model es ima es ixed. In each i e -
a ion, I p edic new mon hly e u ns o he espec i e one-
yea es sample and calcula e he change in ou -o -sample
R2
OOS. I use he a e age a iable impo ance ac oss each an-
nually ained model o ank each ea u e whe e ank one
is he mos impo an cha ac e is ic. To de e mine he ela-
i e impo ance o indi idual a iables o he pe o mance o
each model, I no malize a iable impo ance wi hin a model
o sum o one. I se ing a signal o ze o inc eases he panel
p edic i e ou -o -sample R2
OOS, he a iable impo ance mea-
su e o ha signal is nega i e. The e o e, he no malized
a iable impo ance o o he signals wi hin a model can be
g ea e han 1.
4.3. Machine lea ning po olios
Po olio pe o mance is my p ima y me ic o e alua -
ing he o ecas pe o mance o machine lea ning models. A
he end o each mon h, each model p oduces a p edic ion o
a s ock’s nex mon h’s ela i e e u n ˆ
el
i, +1. Based on hese
o ecas s, I so s ocks om highes o lowes p edic ed e-
u n and assign hem in o decile po olios using NYSE b eak-
poin s.4I eassign and ebalance po olios a he end o
each mon h. I compu e alue-weigh ed e u ns om hold-
ing he decile po olios o e he nex mon h o a oid small
s ocks d i ing he esul s. Finally, I cons uc a ze o-ne in-
es men (long-sho ) po olio ha goes long in he high-
es decile po olio (decile 10) and sho in he lowes decile
po olio (decile 1).
I e alua e he p edic i e pe o mance o h ee di e en
machine lea ning s a egies. The i s s a egy o ms decile
po olios based on he p edic ions o he global neu al ne -
wo k model. The second s a egy o ms decile po olios wi h
sec o -neu al po olio so s based on he global neu al ne -
wo k model p edic ions. This means each sec o ge s indi id-
ual b eakpoin s o he decile so s. The hi d s a egy o ms
decile po olios based on he p edic ions om he sec o -
speci ic neu al ne wo ks. Fi s , I pe o m po olio so s o
each sec o indi idually based on he espec i e model o e-
cas s. Then, I combine he sec o -speci ic decile po olios
in o single decile po olios. Fo example, he op decile po -
olio o mon h con ains all s ocks om he en op decile
po olios ac oss all sec o s in mon h .
4B eakpoin s o he decile so s a e i s de e mined using only s ocks
lis ed on he NYSE. All s ocks a e hen so ed in o decile po olios based
on hese b eakpoin s, ega dless o which exchange hey a e lis ed on. As
he NYSE con ains s ocks wi h la ge a e age ma ke capi aliza ions, his
educes he in luence o small s ocks on he po olio so s.
To compa e he esul s o he h ee neu al ne wo k po -
olio so s, I p o ide each decile po olio’s a e age p edic ed
e u ns, ealized e u ns, and Sha pe a ios. I compu e mean
e u ns and associa ed -s a is ics o he long-sho po olios
o each machine lea ning s a egy. To benchma k he long-
sho e u ns, I conside he adjus ed R2- alue and alphas
om he Capi al Asse P icing Model (CAPM) and Fama and
F ench (2015) i e- ac o model wi h hei associa ed Newey
and Wes (1987) adjus ed -s a is ics using six lags. The ac-
o s a e based on he same sample o 1.3 million mon hly
s ock obse a ions as he neu al ne wo k po olios.
Finally, I benchma k he p edic i e pe o mance o he
sec o -speci ic neu al ne wo ks wi h OLS models in Sec-
ion 5.4. The po olio so s wo k simila ly bu a e based
on p edic ions om sec o -speci ic OLS models. All sec o
models men ioned ou side o Sec ion 5.4 always e e o
sec o -speci ic neu al ne wo ks.
5. Empi ical esul s
5.1. P edic ion pe o mance
The global neu al ne wo k domina es he sec o -speci ic
neu al ne wo ks in ou -o -sample p edic i e pe o mance o
indi idual s ock e u n o ecas s. On he ull sample o e he
ou -o -sample pe iod om Janua y 1994 o Decembe 2022,
he global model achie es a mon hly R2
OOS o 3.37%. The sec-
o models achie e a mon hly R2
OOS o -6.06%, so hey unde -
pe o m a nai e o ecas o ze o o all mon hly s ock e u ns.
Ac oss all en GICS sec o s, he global model ou pe o ms he
espec i e sec o model. Sec o models pe o m pa icula ly
wo se o sec o s wi h only a small sample size.
Table 2compa es he mon hly ou -o -sample s ock-le el
p edic ion pe o mance ac oss all en GICS sec o s be ween
he global model and he en sec o -speci ic models. The sec-
o pe o mance o he global model is de e mined by il-
e ing he ou -o -sample p edic ions o he global neu al ne -
wo k o he espec i e sec o s ocks. In addi ion, Table 2in-
cludes he a e age mon hly obse a ions pe sec o o e he
ull sample pe iod om July 1963 o Decembe 2022. This
demons a es he sample size o each sec o .
The global model p oduces posi i e R2
OOS s a is ics ac oss
all indi idual sec o s. Taking he ull sample o s ocks, he
global model achie es a R2
OOS o 3.37%. The e o e, he p e-
dic ions consis en ly ou pe o m a nai e o ecas o ze o o
all s ocks in all mon hs o e he ou -o -sample pe iod. Excep
o U ili ies, he R2
OOS s a is ics o all sec o s a e la ge han
2%. O e six ou o en sec o s he global model p oduces
R2
OOS abo e 3%, wi h he highes alue a 4.57% o he sec-
o Heal h Ca e. U ili ies appea o be an ou lie wi h 0.11%,
mo e han an o de o magni ude smalle han he R2
OOS o
all o he sec o s. The e is no co ela ion be ween he sec o s’
sample size and he p edic i e pe o mance o he global neu-
al ne wo k. The model achie es a R2
OOS o 3.11% on Com-
munica ion Se ices, he sec o wi h he smalles sample size
and only an a e age o 58 s ocks pe mon h in he sample.
This is mo e han he 2.26% o Financials, he bigges sec o
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581 575
Table 4: S a is ics o neu al ne wo k long-sho po olios
This able summa izes he ou -o -sample s a is ics o he alue-weigh ed long-sho po olios o med om di e en neu al ne wo k model e u n
p edic ions. All s ocks a e so ed in o decile po olios based on hei p edic ed ela i e e u ns (wi h he ma ke componen emo ed) o he nex mon h.
A long-sho po olio buys he highes expec ed e u n s ocks (decile 10) and sells he lowes (decile 1). Resul s a e epo ed o he global model, he
global model wi h sec o -neu al po olio so s, and he sec o models. Panel A p esen s he a e age alue-weigh ed mon hly ull sample mean e u n and
a e age mon hly sub-sample mean e u ns wi h associa ed -s a is ics ( -s a ). Panel B epo s he a e age CAPM alphas and a e age Fama and F ench
(2015) i e- ac o model (FF5) alphas, co esponding Newey and Wes (1987) adjus ed -s a is ics wi h six lags ( -s a α), and co esponding R2. The sample
consis s o US CRSP s ocks, excluding mic ocap s ocks wi h a ma ke capi aliza ion smalle han he 20 h pe cen ile o s ocks lis ed on he NYSE. The
sample uns om Janua y 1994 o Decembe 2022.
Global model Global model +
sec o -neu al so s
Sec o models
Panel A: Pe cen age e u ns
Mean 1994–2022 2.71 2.08 0.99
-s a 11.16 10.36 6.14
Mean 1994–2008 3.24 2.66 0.70
-s a 9.55 9.08 3.06
Mean 2009–2022 2.14 1.46 1.29
-s a 6.24 5.49 5.80
Panel B: Risk-adjus ed pe o mance
CAPM alpha (%) 2.57 2.00 0.98
-s a α8.80 9.55 5.78
R20.03 0.02 -0.003
FF5 alpha (%) 2.44 1.90 0.85
-s a α8.80 8.78 5.98
R20.14 0.12 0.04
mance is signi ican ly mo e subs an ial in he i s hal o he
ou -o -sample pe iod, when i ea ns mo e han ou imes as
much as he sec o neu al ne wo ks.
The global model pe o ms be e om 1994 o 2008
wi h a 3.24% mon hly long-sho e u n and an associa ed
-s a is ic o 9.55 compa ed o a 2.14% e u n om 2009 o
2022 wi h an associa ed -s a is ic o 6.24. This is analogous
o he esul s o Bli z e al. (2023), who ind a weake ou -
o -sample pe o mance o machine lea ning models in hei
la e subsample a e 2004. They base his esul pa ly on
he s eng h o he Size ac o o e di e en ime pe iods.
Size is he mos impo an p edic o o e u ns in my global
neu al ne wo k. Acco ding o Bli z e al. (2023), he s onge
pe o mance o he global model in he ea lie subsample can
be a ibu ed o he excellen pe o mance o he Size ac o
in ea lie pe iods. The Size ac o s a s o pe o m wo se
in he las 20 yea s, which weakens he p o i abili y o he
global neu al ne wo k in he second subsample.
The long-sho po olio based on he so ed p edic ions
o he sec o models unde pe o ms he global model, pa ic-
ula ly in he i s hal o he ou -o -sample pe iod. I achie es
a mon hly mean e u n o 0.7% wi h an associa ed -s a is ic
o 3.06. Al hough s ill less p o i able han he global model,
he sec o -speci ic neu al ne wo ks imp o e in he second
subsample wi h an a e age e u n o 1.29%. Associa ed -
s a is ics a e no a apa , wi h 5.80 o he sec o models
and 6.24 o he global model. The e o e, bo h s a egies a e
simila in isk- e u n p o ile o e he second hal o he ou -
o -sample pe iod.
Two ac o s can d i e he imp o ed pe o mance o he
sec o models in he la e pe iod o he sample. Fi s , no all
sec o neu al ne wo ks ely on Size as an essen ial signal o
he same ex en as he global model. The e o e, hei p o -
i abili y is no as dependen on he pe o mance o he Size
ac o . Second, in he ea lie pa o he sample, he aining
samples o he sec o -speci ic neu al ne wo ks a e iny, wi h
a e y low obse a ions- o-pa ame e s a io. This makes i
mo e challenging o es ima e coe icien s in a neu al ne wo k
wi hou o e i ing on he aining da a.
Table A1 in he Appendix epo s he ou -o -sample pe -
o mance o indi idual GICS sec o long-sho po olios
based on he so ed p edic ions om he en sec o -speci ic
neu al ne wo ks.
Panel B o Table 4summa izes he isk-adjus ed pe o -
mance o he long-sho po olios o each s a egy based
on ac o p icing models. I epo alphas on op o he Cap-
i al Asse P icing Model (CAPM) and he Fama and F ench
(2015) i e- ac o model (FF5) wi h associa ed Newey and

J. Wi e /Junio Managemen Science 10(3) (2025) 561-581576
Wes (1987) adjus ed -s a is ics wi h six lags ( -s a α) and
adjus ed R2wi h espec o each ac o model.
The isk-adjus ed pe o mance yields simila esul s as
he aw long-sho e u ns wi h a supe io global model. All
h ee neu al ne wo k s a egies achie e s a is ically signi i-
can alphas wi h -s a is ics anging om 5.78 o he sec-
o models on he CAPM o 9.55 o he global model wi h
sec o -neu al po olio so s on he CAPM. The global model
p oduces he highes alphas, wi h 2.57% on he CAPM and
2.44% on FF5. The sec o -neu al po olio so s sligh ly
lowe he alphas o he global model o 2.00% on op o he
CAPM and 1.90% on op o FF5. The sec o -speci ic models
span he lowes alphas, wi h 0.98% on he CAPM and 0.85%
on FF5. The CAPM ba ely has any explana o y powe on he
a e age long-sho e u ns o neu al ne wo k o ecas s, wi h
R2ne e exceeding 0.03 o he global model. The i e- ac o
model explains as much as 14% o he a ia ion in he long-
sho po olio based on he global model’s o ecas s. Unsu -
p isingly, he Size ac o is he s a is ically mos signi ican
ac o in eg essions o po olio e u ns on he i e- ac o
model o all neu al ne wo k s a egies.
The esul s o Tables 3and 4a e illus a ed in Figu e 3. I
plo s he cumula i e log e u ns o he alue-weigh ed long
and sho sides o he h ee neu al ne wo k s a egies in he
ou -o -sample pe iod. The long side buys he s ocks in he
highes decile po olio and he sho side sells he lowes
decile po olio. The e o e, e u ns on he sho po olio a e
he ela i e e u ns o he lowes decile s ocks mul iplied by
-1. The cumula i e pe o mance is cu o in Decembe 2008
and es a ed in Janua y 2009 o p esen di e ences in cu-
mula i e e u ns be ween he wo hal es o he ou -o -sample
pe iod.
The global model consis en ly domina es he sec o mod-
els o e ime. Howe e , i s ou pe o mance is mainly in he
i s hal o he ou -o -sample pe iod and ape s o he ea e .
The cumula i e e u ns o all h ee s a egies ollow simila
pa e ns in he second hal o he ou -o -sample pe iod. The
long-sho sp eads a e s ill la ge o he global neu al ne -
wo k a e 2008, bu he magni ude o he e u ns ela i e
o he sec o models is smalle . Sec o -neu al po olio so -
ing p e en s he global model om ou pe o ming he sec o
neu al ne wo ks in he second hal o he ou -o -sample pe-
iod. The pe o mance o long-sho po olios o he global
model is no p edominan ly based on he sho side, which
would aise ques ions abou p ac ical implemen a ion due o
sho ing ic ions.
The e u n se ies o he global model’s long po olio is
s ong ini ially and s a s o shi a e 2000. I s ill deli e s
posi i e esul s, bu wi h highe ola ili y, he o e all magni-
ude o ela i e e u ns is lowe . The s ocks in he op decile
o he global neu al ne wo k’s o ecas s cumula e double he
e u ns in he eigh yea s om 1994 o 2001 han in he ol-
lowing se en yea s. The second hal o he ou -o -sample
pe iod om 2009 o 2023 accumula es oughly he same e-
u ns as he i s eigh yea s. Apa om he shi in he e-
u n se ies du ing he do -com bubble c ash in 2001, global
shocks such as he inancial c isis o 2008 and 2009 o he
COVID-19 pandemic in ea ly 2020 did no cause signi ican
po olio down u ns.
The sho side o he global model gene a es posi i e
e u ns bu unde pe o ms he long po olio in bo h sub-
samples. I s posi i e pe o mance is mainly due o he do -
com bubble c ash om 2000 o 2002. Apa om his pe iod,
cumula i e e u ns inc ease only sligh ly o mo e sideways
o e ex ended pe iods in he plo .
The long and sho sides o he po olios o he global
model wi h sec o -neu al po olios gene ally ollow he
same pa e n. The magni ude o e u ns and he long-sho
sp ead a e smalle . The shi in he e u n se ies a e 2000
is mo e p onounced, and he do -com bubble c ash causes a
po olio down u n.
The op and bo om decile po olios based on he so ed
p edic ions o he sec o neu al ne wo ks do no pe o m well
in he i s hal o he ou -o -sample pe iod. The long po o-
lio gene a es no signi ican e u ns a e 1999 and wipes ou
all accumula ed e u ns in he do -com bubble c ash. The
sho po olio bene i s om his c isis bu o he wise does
no gene a e any subs an ial e u ns. I is he only po olio
o accumula e nega i e e u ns in pa s o he ou -o -sample
pe iod. A e 2008, he cumula i e e u ns o he long and
sho po olios eco e bu ne e each he magni ude o he
global model.
Appendix A2 plo s he cumula i e log e u ns o op and
bo om decile po olios so ed on he ou -o -sample e u n
o ecas s o indi idual sec o -speci ic neu al ne wo k models.
5.4. Benchma king wi h OLS sec o models
Sec o -speci ic neu al ne wo ks unde pe o m a bench-
ma k in he o m o sec o -speci ic o dina y leas squa es
(OLS) models in he ull sample. A long-sho po olio based
on he so ed p edic ions o OLS sec o models gene a es a
highe alue-weigh ed ela i e e u n han he sec o neu-
al ne wo ks. Howe e , he OLS ou pe o mance comes only
om he i s hal o he ou -o -sample pe iod. The ade-
o be ween s able model es ima ion and cap u ing sec o -
speci ic complex in e ac ions is pa icula ly ele an o small
sec o samples. Fo ecas s om he global neu al ne wo k on
he pooled da a ac oss sec o s emain mo e p o i able han
he OLS models.
Sec o models unde pe o m he global model ega ding
s ock-le el o ecas ing and long-sho po olio pe o mance,
e en when sec o -speci ic b eakpoin s so he global model
p edic ions. As a i s s ep owa ds unde s anding he ea-
sons o his obse a ion, I compa e he sec o -speci ic mod-
els wi h ano he machine lea ning echnique. The sec o
models used in his hesis build on a neu al ne wo k a chi-
ec u e wi h h ee hidden laye s (NN3). I compa e hese
wi h sec o -speci ic models based on a simple linea p edic-
i e eg ession model es ima ed by o dina y leas squa es.
The me hodology emains he same; only he machine lea n-
ing model o p edic ing e u ns om s ock-speci ic signals
changes.
Table 5compa es he ou -o -sample pe o mance o long-
sho po olios based on he so ed p edic ions om he wo
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581 577
Figu e 3: Cumula i e pe o mance o op and bo om decile po olios
The igu e plo s he cumula i e log e u ns o op and bo om decile po olios so ed on he ou -o -sample neu al ne wo k e u n o ecas s. Each mon h,
s ocks a e so ed in o alue-weigh ed decile po olios based on he p edic ed ela i e e u ns (wi h he ma ke componen emo ed) om he h ee neu al
ne wo k s a egies. The solid and dash lines ep esen long ( op decile) and sho (bo om decile) posi ions, espec i ely. Fo he sho posi ion, he mon hly
ela i e e u ns o he bo om po olio a e mul iplied by -1. The igu e includes plo s o he cumula i e e u ns o he global model, he global model wi h
sec o -neu al po olio so s, and he sec o models. The sample consis s o US CRSP s ocks, excluding mic ocap s ocks wi h a ma ke capi aliza ion smalle
han he 20 h pe cen ile o s ocks lis ed on he NYSE. The sample uns om Janua y 1994 o Decembe 2022. The cumula i e pe o mance is cu o in
Decembe 2008 and es a ed in Janua y 2009 o p esen di e ences in cumula i e e u ns be ween he wo sub-samples.
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581578
Table 5: Pe o mance o NN3 sec o models e sus OLS sec o models
This able compa es he ou -o -sample pe o mance o he alue-weigh ed long-sho po olios o med om di e en machine lea ning sec o model e u n
p edic ions. Resul s a e epo ed o he NN3 sec o models (same sec o models as in Table 4and Figu e 3) and OLS sec o models. All s ocks a e so ed
in o decile po olios based on hei p edic ed ela i e e u ns (wi h he ma ke componen emo ed) o he nex mon h. A long-sho po olio buys he
highes expec ed e u n s ocks (decile 10) and sells he lowes (decile 1). Panel A compa es he a e age alue-weigh ed mon hly ull sample mean e u n
and a e age mon hly sub-sample mean e u ns wi h associa ed -s a is ics ( -s a ). Panel B isualizes he cumula i e log e u ns o op and bo om decile
po olios so ed on he ou -o -sample e u n o ecas s o he wo machine lea ning s a egies. The solid and dash lines ep esen long ( op decile) and sho
(bo om decile) posi ions, espec i ely. The sample consis s o US CRSP s ocks, excluding mic ocap s ocks wi h a ma ke capi aliza ion smalle han he 20 h
pe cen ile o s ocks lis ed on he NYSE. The sample uns om Janua y 1994 o Decembe 2022.
Panel A: Pe cen age e u ns
Sec o Models Mean
1994-2022
-s a Mean
1994-2008
-s a Mean
2009-2022
-s a
NN3 0.99 6.14 0.70 3.06 1.29 5.80
OLS 1.43 7.80 1.78 6.41 1.05 4.53
Panel B: Cumula i e pe o mance o op and bo om decile po olios
di e en machine lea ning me hods. Panel A epo s a -
e age mon hly alue-weigh ed ou -o -sample e u ns wi h
associa ed -s a is ics ( -s a ). The sec o neu al ne wo ks
(NN3) e u ns a e he same as in Panel A o Table 4. Panel
B isualizes he cumula i e ou -o -sample log e u ns o he
alue-weigh ed long and sho sides o he wo machine
lea ning models.
The OLS sec o models ou pe o m he neu al ne wo k
sec o models o e he whole ou -o -sample pe iod om Jan-
ua y 1994 o Decembe 2022. They achie e a e age mon hly
long-sho po olio e u ns o 1.43% wi h an associa ed -
s a is ic o 7.80. Howe e , he esul s di e o he wo di -
e en subsamples. The ou pe o mance o he OLS models is
based solely on he i s hal o he ou -o -sample pe iod om
1994 o 2008. The OLS sec o models gene a e signi ican ly
highe long-sho e u ns han he NN3 sec o models du ing
his pe iod. The mon hly ela i e e u ns o 1.78% o he
OLS models a e mo e han double he 0.70% achie ed by he
sec o models. The pic u e changes when looking a he la e
subsample om 2009 o 2022. In his pe iod, OLS models
unde pe o m NN3 models, wi h a e age mon hly e u ns o
1.05% and an associa ed -s a is ic o 4.53. Panel B illus a es
hese esul s. The cumula i e log e u ns o he OLS long
and sho po olios emain posi i e h oughou he ou -o -
sample pe iod. They pe o m mo e eliably in he ea ly yea s,
wi h only a iny long-sho sp ead. As wi h all o he machine
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581 579
lea ning models, he gains o he sho po olio a e mainly
due o he do com bubble c ash. The bigges di e ence be-
ween he OLS and NN3 sec o models is he long po olio
in he i s hal o he ou -o -sample pe iod. The cumula i e
e u ns o he op decile s ocks based on OLS o ecas s a e
abou an o de o magni ude highe han hei NN3 coun-
e pa s. In he second subsample om 2009 o 2022, he
e u ns o he OLS long and sho po olios ollow a simila
pa e n o he NN3 e u ns, wi h a sligh unde pe o mance.
These esul s demons a e a ade-o be ween obus
model es ima ion and complex model a chi ec u e o cap-
u e sec o -speci ic nonlinea i ies and a iable in e ac ions.
A simple linea eg ession model such as OLS equi es only
a single pa ame e o each o he 212 signals. The pa-
ame e s o he op imiza ion p oblem a e de i ed om a
closed- o m solu ion. The e o e, OLS models bene i om
a highe a io o obse a ions o pa ame e s and mo e s a-
ble model pa ame e s. Thei ob ious disad an age is he
inabili y o cap u e non-linea i ies and a iable in e ac ions.
The numbe o es ima ed coe icien s in each o my neu al
ne wo k models is 7,489. These numbe s esul in a low
obse a ions- o-pa ame e a io on small da a se s, such as
he indi idual sec o samples, especially a he beginning
o my sampling pe iod when he aining samples a e he
smalles . In low signal- o-noise p oblems like s ock e u n
p edic ion, complex machine lea ning models such as neu al
ne wo ks end o o e i noise a he han ex ac signals.
The NN3 sec o models canno exploi hei ad an age o
being able o cap u e complex sec o -speci ic ela ionships
be ween signals and u u e e u ns. This can lead o he
poo p edic i e pe o mance o neu al ne wo ks o small
sec o s and low po olio e u ns in he i s hal o he ou -
o -sample pe iod. The pooling o da a ac oss sec o s o he
global model imp o es he a io o obse a ions o pa am-
e e s. This explains why i does no su e om he same
p oblems as he sec o al models. The global neu al ne wo k
is s ill able o cap u e nonlinea i ies and a iable in e ac ions
ac oss sec o s and hus signi ican ly ou pe o ms OLS mod-
els in he ull sample, in line wi h p e ious esea ch (e.g.,
Aze edo and Hoegne (2023), Bli z e al. (2023), and Gu
e al. (2020)).
Table A3 in he Appendix summa izes he ou -o -sample
s a is ics o he alue-weigh ed long-sho po olios o med
om di e en sec o -speci ic OLS model e u n p edic ions.
5.5. Sec o alloca ion as a e u n d i e
The global neu al ne wo k demons a es some ou -o -
sample sec o alloca ion powe . In he c oss-sec ion o sec-
o s, i co ec ly p edic s highe ela i e e u ns o he mos
p o i able sec o s and lowe e u ns o he leas p o i able
sec o s. This allows he global model o gene a e highe e-
u ns in he long ( op decile) po olio when po olio so ing
is no sec o -neu al.
As seen in Table 4, sec o -neu al po olio so s wo sen
he p o i abili y o long-sho po olios so ed on he ou -o -
sample e u n o ecas s o he global neu al ne wo k. To be -
e unde s and his di e ence in pe o mance, I b ie ly e alu-
a e he global neu al ne wo k’s po en ial ou -o -sample sec o
alloca ion powe . Table 6summa izes my esul s. Fo each
o he en GICS sec o s, Panel A epo s he a e age alue-
weigh ed ealized mon hly ela i e e u n compa ed o he
a e age p edic ed e u n om he global model. In addi ion,
I epo he a e age mon hly alloca ion o he espec i e sec-
o in he op decile po olio so ed based on o ecas s om
he global model. I ocus on he long po olio, he s onge
d i e o he global model’s p o i abili y han he sho po o-
lio. Panel B plo s he cumula i e alue-weigh ed ela i e log
e u ns pe sec o o e he ou -o -sample pe iod om Jan-
ua y 1994 o Decembe 2022.
In o ma ion Technology ( ech) is he bes -pe o ming sec-
o o e he en i e ou -o -sample pe iod, wi h a mon hly el-
a i e e u n o 0.23%. Heal h Ca e (heal h) ollows in sec-
ond place wi h a e u n o 0.09%, and Ene gy (ene gy) is in
hi d place wi h a e u n o 0.01%, hanks o solid gains a -
e 2020. All o he sec o s gene a e nega i e alue-weigh ed
ela i e e u ns ou -o -sample.5The wo s -pe o ming sec-
o s a e U ili ies (u ili ies), wi h an a e age mon hly e u n o
-0.24%, and Communica ion Se ices (comm), wi h a e u n
o -0.35%. Apa om In o ma ion Technology, he global
model’s e u n p edic ions a e no e y close o he ealized
e u ns, and hey o e s a e expec ed e u ns. S ill, i pe -
o ms ela i ely well in c oss-sec ionally classi ying he sec-
o s in o he co ec ex emes. The global model co ec ly
iden i ies he wo bes -pe o ming sec o s and p edic s U il-
i ies and Communica ion Se ices o be among he h ee
wo s -pe o ming sec o s.
As a esul , he global model shows some (limi ed) ou -
o -sample sec o alloca ion powe . The long po olio based
on global neu al ne wo k o ecas s has high a e age alloca-
ions o In o ma ion Technology (22.93%) and Heal h Ca e
(15.09%). I has an a e age alloca ion o less han 5% o
each o he h ee wo s -pe o ming sec o s: Ma e ials, U il-
i ies, and Communica ion Se ices. The global model gen-
e a es highe e u ns han sec o -neu al po olios by o e -
weigh ing mo e p o i able sec o s and unde weigh ing less
p o i able sec o s in he op decile po olio.
6. Conclusion
I examine he di e ence in p edic i e powe o he c oss-
sec ion o US s ock e u ns be ween a global machine lea n-
ing model and sec o -speci ic models. Based on hei s ong
pe o mance in p e ious esea ch, I use neu al ne wo ks as
he machine lea ning models. The global neu al ne wo k is
ained on he ull sample o s ocks, while he sec o neu-
al ne wo ks a e ained on en di e en GICS sec o s. The
global model consis en ly ou pe o ms he sec o models ou -
o -sample in e ms o p edic i e accu acy and p o i abili y. I
5Equally-weigh ed esul s (no epo ed) show ha almos all sec o s ha e
posi i e ela i e e u ns. On a e age, s ocks wi h smalle ma ke capi al-
iza ion gene a e highe ela i e e u ns. This u he demons a es he
s ong pe o mance o he Size ac o in he sample and jus i ies he high
a iable impo ance o Size in he global neu al ne wo k.
J. Wi e /Junio Managemen Science 10(3) (2025) 561-581580
Table 6: Sec o alloca ion powe o global neu al ne wo k model
This able summa izes pa s o he ou -o -sample sec o alloca ion powe o he global neu al ne wo k model. Panel A compa es he a e age mon hly
ealized ela i e e u ns (wi h he ma ke componen emo ed) pe sec o o he a e age mon hly p edic ed ela i e e u ns om he global model o e he
ou -o -sample pe iod. The sec o s in Panel A a e anked in descending o de o ealized e u ns. Panel A addi ionally epo s he a e age mon hly alloca ion
o each sec o o he long ( op decile) po olio so ed based on he global model’s e u n p edic ions o he nex mon h. Panel B plo s he cumula i e
ela i e log e u ns pe sec o o e he ou -o -sample pe iod. All e u ns a e alue-weigh ed. The ou -o -sample pe iod uns om Janua y 1994 o Decembe
2022. The sample consis s o US CRSP s ocks, excluding mic ocap s ocks wi h a ma ke capi aliza ion smalle han he 20 h pe cen ile o s ocks lis ed on he
NYSE.
Panel A: Pe cen age e u ns and long po olio alloca ion
Sec o Realized e u ns P edic ed e u ns Long po olio alloca ion
In o ma ion Technology 0.23 0.26 22.93
Heal h Ca e 0.09 0.17 15.09
Ene gy 0.01 0.06 7.21
Consume Disc e iona y -0.10 0.14 14.52
Consume S aples -0.11 -0.03 3.74
Indus ials -0.11 0.12 11.54
Financials -0.13 0.14 12.96
Ma e ials -0.16 0.11 4.50
U ili ies -0.24 -0.02 3.96
Communica ion Se ices -0.35 0.05 4.36
Panel B: Cumula i e ela i e log e u ns pe sec o
de i es mos o i s p edic i e powe om Size as an inpu
signal. A long-sho po olio based on he so ed p edic-
ions o he global model gene a es signi ican e u ns and
Sha pe a ios o e he en i e ou -o -sample pe iod. The sec-
o models gene a e nega i e ou -o -sample R2
OOS and hei
long-sho po olio e u ns a e lowe , especially in he ea ly
ou -o -sample pe iod. Complex models such as non-linea
neu al ne wo ks s uggle o exploi hei ad an ages on small
sec o -speci ic samples and unde pe o m simple OLS mod-
els. The esul s o he global model suppo he ecen li e -
a u e on he s ong p edic i e powe o neu al ne wo ks o
he c oss-sec ion o s ock e u ns.
Re e ences
Amihud, Y. (2002). Illiquidi y and s ock e u ns: c oss-sec ion and ime-
se ies e ec s. Jou nal o Financial Ma ke s,5(1), 31–56.
A amo , D., Cheng, S., & Me zke , L. (2023). Machine Lea ning s. Eco-
nomic Res ic ions: E idence om S ock Re u n P edic abili y.
Managemen Science,69(5), 2587–2619.
Aze edo, V., & Hoegne , C. (2023). Enhancing s ock ma ke anomalies wi h
machine lea ning. Re iew o Quan i a i e Finance and Accoun ing,
60(1), 195–230.
Aze edo, V., Hoegne , C., & Veliko , M. (2024). The Expec ed Re u ns on
Machine-Lea ning S a egies. SSRN Elec onic Jou nal.
Aze edo, V., Kaise , G. S., & Muelle , S. (2023). S ock ma ke anomalies and
machine lea ning ac oss he globe. Jou nal o Asse Managemen ,
24(5), 419–441.
Bli z, D., Hanaue , M. X., Hoog eijling, T., & Howa d, C. (2023). The Te m
S uc u e o Machine Lea ning Alpha. SSRN Elec onic Jou nal, 1–
40.

J. Wi e /Junio Managemen Science 10(3) (2025) 561-581 581
Cakici, N., Fiebe g, C., Me ko, D., & Za emba, A. (2023). Machine lea ning
goes global: C oss-sec ional e u n p edic abili y in in e na ional
s ock ma ke s. Jou nal o Economic Dynamics and Con ol,155,
1–32.
Ca aglia, S. M. F. G., Se on, J., Scowc o , A., & Smi h, B. (2006). Global
S yle In es ing. The Jou nal o Po olio Managemen ,32(4), 10–
22.
Chen, A. Y., & Zimme mann, T. (2022). Open Sou ce C oss-Sec ional Asse
P icing. C i ical Finance Re iew,27(2), 207–264.
Chen, L., Pelge , M., & Zhu, J. (2024). Deep Lea ning in Asse P icing. Man-
agemen Science,70(2), 714–750.
Cine , C. (2019). Do indus y e u ns p edic he s ock ma ke ? A ep ise us-
ing he andom o es . Qua e ly Re iew o Economics and Finance,
72, 152–158.
D obe z, W., & O o, T. (2021). Empi ical asse p icing ia machine lea n-
ing: e idence om he Eu opean s ock ma ke . Jou nal o Asse
Managemen ,22(7), 507–538.
Fama, E. F., & F ench, K. R. (2015). A i e- ac o asse p icing model. Jou nal
o Financial Economics,116(1), 1–22.
F ench, K. R. (2024). Cu en Resea ch Re u ns. Re ie ed No embe 1,
2024, om h p://mba. uck.da mou h.edu/pages/ acul y/k
en. ench/da a_lib a y.h ml
F eybe ge , J., Neuhie l, A., & Webe , M. (2020). Dissec ing Cha ac e is ics
Nonpa ame ically. The Re iew o Financial S udies,33(5), 2326–
2377.
Gu, S., Kelly, B., & Xiu, D. (2020). Empi ical Asse P icing ia Machine Lea n-
ing. The Re iew o Financial S udies,33(5), 2223–2273.
Hanaue , M. X., & Kalsbach, T. (2023). Machine lea ning and he c oss-
sec ion o eme ging ma ke s ock e u ns. Eme ging Ma ke s Re-
iew,55, 101022.
Hou, K., Xue, C., & Zhang, L. (2020). Replica ing Anomalies. The Re iew o
Financial S udies,33(5), 2019–2133.
Kim, K. J., Lee, C., & Ti as, S. L. (2013). The e ec s o adjus ing he esid-
ual income model o indus y and i m-speci ic ac o s when p e-
dic ing u u e abno mal e u ns. Asia-Paci ic Jou nal o Financial
S udies,42(3), 373–402.
Leippold, M., Wang, Q., & Zhou, W. (2022). Machine lea ning in he Chinese
s ock ma ke . Jou nal o Financial Economics,145(2), 64–82.
Lewellen, J. (2015). The C oss-sec ion o Expec ed S ock Re u ns. C i ical
Finance Re iew,4(1), 1–44.
Liu, X., Pong, E. S. Y., Shackle on, M. B., & Zhang, Y. (2014). Op ion-Implied
Vola ili ies and S ock Re u ns: E idence om Indus y-Neu al
Po olios. The Jou nal o Po olio Managemen ,41(1), 65–77.
Mo i z, B., & Zimme mann, T. (2016). T ee-Based Condi ional Po olio
So s: The Rela ion be ween Pas and Fu u e S ock Re u ns.
SSRN Elec onic Jou nal, 1–81.
MSCI Inc. (2024). The Global Indus y Classi ica ion S anda d (GICS®). Re-
ie ed No embe 1, 2024, om h ps://www.msci.com/ou -sol
u ions/indexes/gics
Nagel, S. (2005). Sho sales, ins i u ional in es o s and he c oss-sec ion o
s ock e u ns. Jou nal o Financial Economics,78(2), 277–309.
Newey, W. K., & Wes , K. D. (1987). A Simple, Posi i e Semi-De ini e, He -
e oskedas ici y and Au oco ela ion Consis en Co a iance Ma-
ix. The Econome ic Socie y,55(3), 703–708.
Rapach, D. E., S auss, J. K., Tu, J., & Zhou, G. (2019). Indus y Re u n
P edic abili y: A Machine Lea ning App oach. The Jou nal o Fi-
nancial Da a Science,1(3), 9–28.
Rasekhscha e, K. C., & Jones, R. C. (2019). Machine Lea ning o S ock
Selec ion. Financial Analys s Jou nal,75(3), 70–88.
Tobek, O., & H onec, M. (2021). Does i pay o ollow anomalies esea ch?
Machine lea ning app oach wi h in e na ional e idence. Jou nal
o Financial Ma ke s,56.
U.S. Depa men o Labo . (2025). S anda d Indus ial Classi ica ion (SIC)
Manual. Re ie ed No embe 1, 2024, om h ps://www.osha.g
o /da a/sic-manual