Kaya, De imi; Reichmann, Do on; Reichmann, Milan
A icle — Published Ve sion
Ou ‐o ‐sample p edic abili y o i m‐speci ic s ock p ice
c ashes: A machine lea ning app oach
Jou nal o Business Finance & Accoun ing
P o ided in Coope a ion wi h:
John Wiley & Sons
Sugges ed Ci a ion: Kaya, De imi; Reichmann, Do on; Reichmann, Milan (2024) : Ou ‐o ‐sample
p edic abili y o i m‐speci ic s ock p ice c ashes: A machine lea ning app oach, Jou nal o Business
Finance & Accoun ing, ISSN 1468-5957, Wiley, Hoboken, NJ, Vol. 52, Iss. 2, pp. 1095-1115,
h ps://doi.o g/10.1111/jb a.12831
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/319309
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h p://c ea i ecommons.o g/licenses/by/4.0/
Recei ed: 13 Ma ch 2023 Re ised: 8 Augus 2024 Accep ed: 12 Augus 2024
DOI: 10.1111/jb a.12831
ARTICLE
Ou -o -sample p edic abili y o i m-speci ic s ock
p ice c ashes: A machine lea ning app oach
De imi Kaya1Do on Reichmann2Milan Reichmann3
1F ied ich-Alexande -Uni e si y o
E langen-Nü nbe g, Chai o Business
Analy ics and Sus ainabili y, Nü nbe g,
Ge many
2Accoun ing Depa men , Goe he Uni e si y
F ank u , F ank u , Ge many
3Chai o Banking and Finance, Leipzig
Uni e si y, Leipzig, Ge many
Co espondence
Do on Reichmann, Theodo - W.-Ado no-Pla z
1, 60629 F ank u am Main, Ge many.
Email: [email p o ec ed]
Abs ac
We use machine lea ning me hods o p edic i m-speci ic
s ock p ice c ashes and e alua e he ou -o -sample p edic-
ion pe o mance o a ious me hods, compa ed o adi-
ional eg ession app oaches. Using inancial and ex ual
da a om 10-K ilings, ou esul s show ha a logis ic eg es-
sion wi h inancial da a inpu s pe o ms easonably well
and some imes ou pe o ms newe classi ie s such as an-
dom o es s and neu al ne wo ks. Howe e , we ind ha a
s ochas ic g adien boos ing model sys ema ically ou pe -
o ms he logis ic eg ession, and o ecas s using sui able
combina ions o inancial and ex ual da a inpu s yield signi -
ican ly highe p edic ion pe o mance. O e all, he e idence
sugges s ha machine lea ning me hods can help p edic
s ock p ice c ashes.
KEYWORDS
machine lea ning, na u al language p ocessing, s ock p ice c ash
isk, ex ual disclosu es
1INTRODUCTION
S ock p ice c ashes a e p e alen in in e na ional inancial ma ke s (An e al., 2018;Jin&Mye s,2006). S ud-
ies ex ensi ely show ha economic ac o s, such as inancial opaci y, agency cos s and manage ial incen i es,
can help explain s ock p ice c ashes (Hong e al., 2017; Hu on e al., 2009; Kim e al., 2019; Kim e al.,
This is an open access a icle unde he e ms o he C ea i e Commons A ibu ion License, which pe mi s use, dis ibu ion and
ep oduc ion in any medium, p o ided he o iginal wo k is p ope ly ci ed.
© 2024 The Au ho (s). Jou nal o Business Finance & Accoun ing published by John Wiley & Sons L d.
J Bus Fin Acc. 2025;52:1095–1115. wileyonlinelib a y.com/jou nal/jb a 1095
1096 KAYA ET AL.
2011a).1Mos s udies ela e economic de e minan s o s ock p ice c ash isks using wi hin-sample analyses. Howe e ,
in e ences abou he p edic abili y o s ock p ice c ashes a e s ill limi ed. I in es o s ail o de ec po en ial h ea s,
s ock p ices can de ia e om hei undamen al alues, inc easing he isk o u u e p ice c ashes. Hence, me hods
ha help o iden i y c ash-p one i ms would o e signi ican alue o in es o s.
We use machine lea ning me hods o p edic i m-speci ic s ock p ice c ashes.2Speci ically, we e alua e he ou -o -
sample p edic ionpe o manceo machinelea ning,compa ed o adi ional eg essionapp oaches.While esea chon
machine lea ning in accoun ing and inance ypically ocuses on nume ical da a as p edic o s (Bali e al., 2023;Chen
e al., 2022; Gu e al., 2020), we also conside ex ual disclosu es as an in eg al pa o he inancial epo ing package.
Fo ins ance, Lewis and Young (2019) documen a subs an ial inc ease in ex ual i m disclosu es o e ime. Machine
lea ning me hods can unco e complex pa e ns in bo h inancial and ex ual da a ha help p edic i m ou comes
(e.g., Be omeu e al., 2021; Bochkay e al., 2023; El-Haj e al., 2020). Hence, ou s udy aims o le e age hese empi ical
me hods and es hei pe o mance o ou -o -sample s ock p ice c ash p edic ions.
To conduc ou analyses, we collec a sample o 39,583 US i m-yea obse a ions om he pe iod 1996 o 2018
and calcula e a wide se o 37 inancial da a inpu s. We u he e ie e he Managemen Discussion and Analysis
(MD&A) sec ions o 10-K ilings om he SEC’s online EDGAR sys em o build ex ual inpu s because esea ch inds
ha ex ual disclosu es o 10-K ilings con ain p edic i e signals o s ock p ice c ashes (e.g., E ug ul e al., 2017;Kim
e al., 2019; Reichmann, 2023).
As a benchma k model, we use a logis ic eg ession (LOGIT) wi h nume ical inancial da a inpu s. This adi ional
eg ession app oach is commonly used in bo h esea ch and p ac ice (Bu a u e al., 2016; Jones & Henshe , 2004,
2007). In addi ion, we conside he suppo ec o machine (SVM; Vapnik, 1998), wo models based on decision ees,
namely, he andom o es (RF; B eiman, 2001) and he s ochas ic g adien boos ing (SGB; F iedman, 2002), and wo
neu al ne wo ks, a dense neu al ne wo k o nume ical inpu s (NN) and a con olu ional neu al ne wo k o ex ual
inpu s (CNN) (e.g., Hin on e al., 2006; LeCun e al., 1989, 2015).
While nume als can be di ec ly used as inpu s o a machine lea ning model, ex ep esen s na u al language, which
equi es con e sion o nume ical ep esen a ions. Ou empi ical app oach is as ollows. Fi s , we mimic p e ious s ud-
ies on c ash isk. Speci ically, we use simple con en analyses o es ima e high-le el ex ual cha ac e is ics ha cap u e
he one, ambigui y and complexi y o MD&A and use hose as inpu s o a LOGIT (E ug ul e al., 2017; Kim e al., 2019).
Second, o p o ide ou model wi h mo e de ailed ex ual in o ma ion, we es ima e nume ical documen ep esen a-
ions using e m equency-in e se documen equency (TF-IDF) weigh s.3This app oach yields la ge ec o s ha
con ain ela i e e m equencies o each MD&A, which can be used as inpu s o a LOGIT, SVM, RF and SGB. Finally,
we conside a wo d embedding app oach. Wo d embeddings a e ec o ep esen a ions o wo ds and ph ases ha
cap u e seman ic in o ma ion. We use wo d2 ec o gene a e wo d embeddings (Mikolo , Chen, e al., 2013), which
se e as inpu s o a CNN. This is ou only model capable o conside ing sequen ial dependencies in inancial na a i es.
To ain ou models, we employ a olling sample spli ing scheme (e.g., Chen e al., 2022;Gue al.,2020). Fo each
yea in ou es pe iod, 2001–2018, we use he i e p eceding yea s o ain, alida e and es ima e he models ha a e
hen es ed in nex yea ’s hold-ou -sample. While his app oach equi es us o eshly ain new models e e y yea in
he es pe iod, i allows he models o be con inuously ained on ecen da a (Chen e al., 2022). Mo eo e , we ecode
se ial s ock p ice c ashes o i ms ha span bo h he model es ima ion and es se (hence o h: se ial c ashes) as non-
c ash obse a ion in he aining se . This app oach mi iga es conce ns ha lexible machine lea ning models iden i y
c ash-p one i ms a he han c ash-p one i m-yea s. Consis en wi h ecen wo k (e.g., Be omeu e al., 2021;Chen
e al., 2022; Mai e al., 2019), we e alua e ou models using he a ea unde he ecei e ope a ing cha ac e is ic (ROC)
cu e (AUC) and a se o ca ch a es wi h a ying p obabili y cu -o s.
1The li e a u e de ines s ock p ice c ashes as ex emenega i e ou lie s in he dis ibu ion o i m e u ns (e.g., Hu on e al., 2009;Kime al.,2019; Kim e al.,
2011a, 2011b).
2The e ha e been se e al ecen s udies using machine lea ning in accoun ing and inance, including in he a ea o ea nings p edic ion (Chen e al., 2022;
Jones e al., 2023), accoun ing miss a emen s (Be omeu e al., 2021) and dis ess p edic ion (Jones e al., 2015, 2017).
3See B own and Tucke (2011) o an explana ion o TF-IDF weigh s.
KAYA ET AL.1097
Ou esul s sugges ha a LOGIT wi h nume ical inpu s (hence o h: LOGIT(Num)) pe o ms easonably well and
se es as a s ong benchma k (Jones e al., 2015, 2017). The model yields an AUC o 55.30%, signi ican ly highe
han he 50% o a andom guess. Mo eo e , i ou pe o ms he SVM(Num), RF(Num) and NN(Num). Howe e , he
SGB(Num) yields an AUC o 56.26%, which signi ican ly exceeds he AUC o LOGIT(Num), sugges ing ha machine
lea ning me hods can help imp o e he p edic ion o s ock p ice c ashes. The SGB is also he bes -pe o ming model
wi h ex ual inpu s om 10-K ilings. We ind ha an SGB ha uses TF-IDF weigh s, SGB(Tex ), yields an AUC o
56.18%, ou pe o ming he nume ical benchma k LOGIT(Num). Mo eo e , using a LOGIT model wi h a se o high-
le el ex ual cha ac e is ics as commonly employed in wi hin-sample s udies on c ash isk signi ican ly unde pe o ms
SGB(Tex ) (E ug ul e al., 2017; Kim e al., 2019). This inding suppo s he iew ha machine lea ning is powe ul o
de ec p edic i e pa e ns ha a e di icul o unco e wi h simple con en analyses (e.g., Bochkay e al., 2023;El-Haj
e al., 2019).
We also es whe he combining nume ical and ex ual da a inpu s u he imp o es c ash p edic ions. The e o e,
we es wo app oaches. Fi s , we conca ena e nume ical and ex ual inpu s, esul ing in a new inpu ec o wi h com-
bined da a inpu s. We hen e ain ou models using hese combined inpu s. Second, we employ an a e age ensemble
app oach by combining he p obabili y es ima es o wo sepa a e models using a simple a e age. We ind ha he
easy- o-implemen a e age ensemble app oach pe o ms be e . Combining he p obabili y es ima es o SGB(Num)
and SGB(Tex ) using a simple a e age leads o a signi ican imp o emen in p edic i e powe o e using SGB(Num).
This esul sugges s ha ex ual inpu s con ain in o ma ion ha is inc emen ally in o ma i e o nume ical inancial
da a o ou -o -sample c ash p edic ions. An analysis o he model pe o mance o e ime e eals ha his inding is
sys ema ic du ing ou es pe iod, exis s du ing almos all es yea s and does no diminish o e ime.
We hen examine he inne wo kings o ou models by iden i ying meaning ul impo an p edic o s o bo h
SGB(Num) and SGB(Tex ). Speci ically, we pe o m pe mu a ion ea u e impo ance, a model inspec ion echnique
ha obse es he deg ada ion o a model’s p edic i e powe when andomly shu ling alues o a single p edic o .
We ind ha common nume ical inancial a iables used in he li e a u e, such as i m size, he nega i e skewness
o e u ns, e u n on asse s and book- o-ma ke a ios help p edic s ock p ice c ashes (e.g., Hu on e al., 2009;Kim
e al., 2019). Tu ning o ex ual p edic o s, we ind ha e ms ela ed o i m g ow h such as “g ow h,” “acquisi ion” and
“spending” as well as e ms signaling h ea s such as “downg ading” and “may ne e succeed” help SGB(Tex ) dis in-
guish c ash om non-c ash obse a ions. These p edic o s a e gene ally consis en wi h he li e a u e (Hu on e al.,
2009; Reichmann, 2023).
Finally, we es he sensi i i y o ou models o ecoding se ial c ashes as non-c ash obse a ions. Failing o co -
ec o se ial c ashes can lead o in la ed model pe o mance i models end o iden i y c ash-p one i ms a he
han c ash-p one i m-yea s. We p o ide e idence ha ailing o co ec o se ial c ashes subs an ially in la es he
model pe o mance o a ious model a chi ec u es. Ou indings should cau ion u u e esea ch when es ing he
ou -o -sample p edic abili y o s ock p ice c ashes.
This s udy con ibu es o p io li e a u e in wo ways. Fi s , i ex ends he b oad li e a u e on s ock p ice c ash isk.
A dominan heme in he li e a u e is ha economic ac o s, such as inancial opaci y, agency cos s o manage ial incen-
i es help explain s ock p ice c ashes (e.g., Hong e al., 2017; Hu on e al., 2009; Kim e al., 2011a), bu mos s udies
ocus on wi hin-sample analyses o iden i y he de e minan s o c ash isk. We con ibu e o he li e a u e by examin-
ing he ou -o -sample p edic abili y o s ock p ice c ashes using machine lea ning me hods. Fu he , ou esul s d aw
a en ion o he high p edic ion powe o ex ual da a based on MD&A disclosu es. Sui able combina ions o inan-
cial and ex ual da a inpu s yield signi ican ly highe p edic ion pe o mance, a inding ha should ma e o in es o s
and egula o s. Fo ins ance, he SEC and o he egula o s aim o pu mo e emphasis on i ms’ na a i e disclosu es
(Eaglesham, 2013). Ou esul s sugges ha machine lea ning me hods can help unco e signals om disclosu es ha
a e associa ed wi h c ash isk.
Second, we complemen he li e a u e in accoun ing and inance using machine lea ning wi h big da a in inan-
cial ma ke s (e.g., Bianchi e al., 2020; Chen e al., 2022; Gu e al., 2020; Jones e al., 2023). Fo example, Gu e al.
(2020) apply machine lea ning algo i hms o s udy he beha io o expec ed s ock e u ns. Chen e al. (2022)use
1098 KAYA ET AL.
TABLE 1 Sample selec ion.
Da a il e s Fi m-yea s Fi ms
Ac i e non inancial/nonu ili y US i ms on Re ini i (1996–2018) 85,584 8,026
Fiscal yea -end p ice ≥1$ 70,780 7,794
≥26 weekly e u ns pe yea 64,258 6,783
Managemen Discussion and Analysis (MD&A) da a om EDGAR 39,583 3,563
No e: This able p esen s he da a il e s o ou sample selec ion.
high-dimensional inancial da a o p edic he di ec ion o one-yea -ahead ea nings changes. We ex end his esea ch
by examining whe he machine lea ning can de ec p edic i e signals in bo h inancial and ex ual da a ha help p e-
dic s ock p ice c ashes. We ind ha an SGB sys ema ically ou pe o ms o he algo i hms in p edic ing s ock p ice
c ashes. Fu he , ou esul s sugges ha he combina ion o nume ical and ex ual inpu s can imp o e model accu acy.
While he li e a u e on s ock p ice c ashes p edominan ly ocuses on US capi al ma ke s, he use and impo ance
o machine lea ning o p edic s ock p ice c ashes in in e na ional inancial ma ke s is likely o g ow in he nea u u e.
S a ing wi h 2027, ce ain inancial and sus ainabili y- ela ed in o ma ion o Eu opean lis ed i ms will be iled ia he
cen alized Eu opean Single Access Poin (ESAP), which is likely o os e he access and p ocessing o i m disclosu es
(e.g., El-Haj e al., 2019; Kaya & Seebeck, 2019). Fu he , da a collec ion cos s a e likely o be lowe h ough a cen al-
ized documen deposi o y, such as ESAP, whe e annual epo s will be accessible in a s uc u ed, elec onic epo ing
o ma (e.g., he Eu opean Single Elec onic Fo ma ).4
2SAMPLE SELECTION AND MODEL INPUTS
2.1 Sample selec ion
Ou ini ial sample consis s o non inancial and nonu ili y US lis ed i ms a ailable on Re ini i Da as eam and Re ini i
Wo ldscope (Reichmann & Reichmann, 2022).5The sample s a s in 1996 when all publicly lis ed US i ms had o ile
inancial epo s elec onically and ends in 2018. We d op obse a ions wi h a iscal yea -end sha e p ice below $1
and obse a ions wi h less han 26 a ailable weekly e u ns in a iscal yea . Fo each i m, we collec a ailable 10-K
ilings om he SEC’s EDGAR sys em and combine he iles wi h he sample.6We ocus on he MD&A sec ions o 10-K
ilings, which p io li e a u e inds o be associa ed wi h u u e c ash isk (Reichmann, 2023; Reichmann e al., 2022).
The sample selec ion is summa ized in Table 1. Ou inal sample consis s o 39,583 i m-yea obse a ions o 3,563
unique i ms.
2.2 S ock p ice c ashes
We ollow p io li e a u e o measu e s ock p ice c ashes and calcula e i m-speci ic weekly e u ns (Wj, ) by es ima -
ing he ollowing expanded index model o each i m-yea (Hu on e al., 2009; Kim e al., 2019; Kim e al., 2011b;
Reichmann & Reichmann, 2022):
j,𝜏=𝛽0+𝛽1 m,𝜏−1+𝛽2 i,𝜏−1+𝛽3 m,𝜏+𝛽4 i,𝜏+𝛽5 m,𝜏+1+𝛽6 i,𝜏+1+𝜖j,𝜏.
4We e e in e es ed eade s o El-Haj e al. (2019).
5We apply he sc eens and il e s p oposed by Schmid e al. (2019) when cons uc ing ou ini ial sample.
6We use he e m “10-K” o e e o he o m 10-K and i s a ian s: 10-K405, 10KSB and 10KSB40.
KAYA ET AL.1099
j,𝜏deno es s ock e u ns o i m jin week 𝜏. m,𝜏is he weekly e u n on he CRSP alue-weigh ed ma ke index mand
i,𝜏deno es he weekly e u n on he Fama–F ench alue-weigh ed index o indus y i. To a oid look-ahead biases, we
de ine a iscal yea as he 12-mon h pe iod ending h ee mon hs a e he iscal yea -end o accoun o he epo ing
lag o he 10-K (e.g., Kim e al., 2019; Reichmann e al., 2022). The i m-speci ic weekly e u n Wj,𝜏is calcula ed as
he na u al loga i hm o 1 plus he eg ession esidual 𝜖j,𝜏. Finally, he ou come CRASH +1is an indica o a iable ha
equals 1 i a i m-speci ic weekly e u n Wj,𝜏d ops 3.09 s anda d de ia ions below i s yea ly mean in he pe iod +1
and 0 o he wise (Reichmann & Reichmann, 2022). We chose 3.09 s anda d de ia ions o gene a e a 0.1% equency in
he log-no mal dis ibu ion (e.g., Hu on e al., 2009; Kim e al., 2019; Kim e al., 2011a, 2011b).
2.3 Nume ical inpu s
We compile a lis o inancial a iables ha a e associa ed wi h c ash isk (e.g., Chen e al., 2001; Hu on e al., 2009;
Kim e al., 2019;Wu&Lai,2020). Because he lexibili y o machine lea ning me hods enables conside ing a b oade
inpu se compa ed o adi ional econome ic echniques, we u he include inancial a ios ha a e in luen ial p e-
dic o s in ela ed asks such as bank up cy p edic ion (e.g., Mai e al., 2019; Reichmann & Reichmann, 2022). In o al,
we collec a se o 37 nume ical inancial a iables. Table 2, panel A, p o ides de ini ions o ou nume ical inpu s. All
nume ical inpu s a e winso ized a he 1% and 99% le els o a oid ex eme ou lie s. We impu e missing alues wi h
ze os.
2.4 Tex ual inpu s
We also examine he p edic i e powe o ex ual da a, which canno be di ec ly ed in o a machine lea ning model.
Hence, we con e ex o nume ical ep esen a ions ha cap u e he con en o a gi en documen . We conside h ee
di e en app oaches wi h an inc easing deg ee o complexi y (Reichmann & Reichmann, 2022).
Fi s , we es ima e ex ual cha ac e is ics using a simple con en analysis. This app oach con e s ex in o a sco e
ha ep esen s a ex ual cha ac e is ic such as one o linguis ic complexi y. Following p e ious esea ch, we es i-
ma e ex ual cha ac e is ics ha a e associa ed wi h u u e c ash isk. Fo ins ance, p e ious esea ch inds ha i ms
wi h mo e ambiguous and complex inancial epo s a e p one o s ock p ice c ashes (e.g., E ug ul e al., 2017;Kim
e al., 2019). The e o e, we calcula e he ac ion o ambiguous wo ds using he Fin-Unc (UNCERTAIN)andMW-Weak
(WEAK_MODAL) wo d lis s o Lough an and McDonald (2011) and he (modi ied) FOG index p oposed by Kim e al.
(2019) as ex ual p edic o s (FOG and MODFOG).
In addi ion, we p oxy o he one o he MD&A using he Fin-Neg wo d lis (NEGATIVE) o Lough an and McDonald
(2011) because disclosu e one is likely o be in o ma i e abou c ash isk (e.g., Fu e al., 2021; Reichmann, 2023).
Finally, we use p oxies o in o ma ion quan i y by calcula ing he na u al loga i hm o 1 plus he o al numbe o wo ds
inan MD&A (LOGLENGTH) and i s ile size in megaby es (LOGFILESIZE; e.g., E ug ul e al., 2017; Lough an & McDonald,
2014). Table 2, panel B, p o ides de ini ions o ex ual MD&A cha ac e is ics.
Second, we es ima e mo e de ailed documen ep esen a ions by compu ing he e m TF-IDF weigh s o wo ds
and ph ases in each MD&A as ollows (Reichmann & Reichmann, 2022):
wi,j =⎧
⎪
⎨
⎪
⎩
i,jlog (N
d i),i i,j ≥1
0 o he wise.
Ndeno es he o al numbe o documen s in a sample; d iis he numbe o documen s con aining he e m i;and
i,j p esen s he wo d coun o e m iin documen j.TheTF i,j measu es he impo ance o a e m wi hin a docu-
men , whe eas he IDF log( N
d i
) adjus s o he equency o a e m in he en i e sample o documen s (see B own &
1100 KAYA ET AL.
TABLE 2 Va iable de ini ions.
P edic o De ini ion P edic o De ini ion
Panel A: Nume ical p edic o s
ACTLCT Cu en asse s/ o al liabili ies LCTSALE Cu en liabili ies/sales
ADJROTA Indus y-adjus ed e u n on angible
asse s
LEV To al liabili ies/ o al asse s
APSALE Accoun s payable/sales LOGAT Log( o al asse s)
CASHAT Cash and sho - e m in es men s/ o al
asse s
LOGMV Log(ma ke alue)
CHAT Cash/ o al asse s LOGSALE Log(sales)
CHLCT Cash/cu en liabili ies MTB Ma ke alue/book alue
CFVOL S anda d de ia ion o cash low/ o al
asse s o e he i e p eceding iscal yea s
NCSKEW Nega i e skewness o i m-speci ic
weekly s ock e u ns
DTURN De ended mon hly u no e NIAT Ne income/ o al asse s
EARNVOL S anda d de ia ion o EBIT/ o al asse s
o e he i e p eceding iscal yea s
NISALE Ne income/sales
EBITAT EBIT/ o al asse s OPAQUE Th ee-yea s mo ing sum o
disc e iona y acc uals
EBITDAAT EBITDA/ o al asse s RETAT Re ained ea nings/ o al asse s
EBITSALE EBIT/sales RELCT Re ained ea nings/cu en liabili ies
FAT To al deb / o al asse s ROA Ope a ing income/ o al asse s
HHI He indahl–Hi schman Index ROS Ope a ing income/sales
INVCHINVT G ow h o in en o ies/in en o ies ROTA Re u n on angible asse s
LCTAT Cu en liabili ies/ o al asse s SALESVOL S anda d de ia ion o sales/ o al
asse s o e he i e p eceding iscal
yea s
LCTCHAT (Cu en liabili ies—cash)/ o al asse s SIGMA S anda d de ia ion o weekly
i m-speci ic s ock e u ns
LCTLT Cu en liabili ies/ o al liabili ies
Panel B: Tex ual MD&A cha ac e is ics
LOGLENGTH Log(numbe o wo ds) NEGATIVE Nega i e wo ds/ o al wo ds
WEAK_MODAL Weak modal wo ds/ o al wo ds UNCERTAIN Unce ain y wo ds/ o al wo ds
FOG (wo ds pe sen ence +pe cen age o
complex wo ds) ×0.4
MODFOG (wo ds pe sen ence +pe cen age o
modi ied complex wo ds) ×0.4
LOGFILESIZE Log( ilesize in megaby es)
No e: This able p esen s a iable de ini ions. Panel A p o ides he de ini ions o 37 inancial a ios employed in ou nume ical
models. Panel B p o ides he de ini ions o ex ual MD&A cha ac e is ics.
Tucke , 2011, o de ails). Gi en a ocabula y o size |V|, a documen jis ep esen ed as a ec o wi h |V|dimen-
sions. Each dimension ep esen s a wo d in he ocabula y and i s co esponding TF-IDF weigh , wi,j. Because ou
machine lea ning models equi e inpu s o ha e he same leng hs, we conside only he 40,000 mos common e ms
in ou sample o MD&A.7Imposing an uppe limi o he dimensionali y o documen ec o s is common in he
7Fo ou main es s, we employ a olling sample spli ing scheme. Hence, o each olling spli , we es ima e new TF-IDF ep esen a ions o he sample o
documen s in a gi en olling spli .
KAYA ET AL.1101
li e a u e because i ensu es ha he dimensionali y o documen ec o s is no d i en by ex eme ou lie s (e.g.,
Mai e al., 2019).8
Thi d, we conside a wo d embedding app oach. Wo d embeddings a e a ype o wo d ep esen a ion ha allows
wo ds o be ep esen ed as ec o s in a ec o space. These ec o ep esen a ions a e de i ed h ough con ex ual
analysis o wo d occu ences wi hin he co pus o MD&A, enabling he cap u e o seman ic ela ionships be ween
wo ds. Fo ins ance, wo ds ha a e seman ically simila end o be close in he ec o space. To cons uc wo d
embeddings o ou sample o MD&A, we ollow p e ious li e a u e (e.g., Du e al., 2022; Li e al., 2021; Mai e al., 2019)
and use wo d2 ec (Mikolo , Chen, e al., 2013). We es ima e 50-, 100-, 200- and 300-dimensional wo d embeddings
o wo ds occu ing in ou sample o MD&A.9
Fo bo h he es ima ions o TF-IDF weigh s and wo d embeddings, ex ensi e ex p ep ocessing is pe o med o
educe ea u e dimensionali y, imp o e gene aliza ion and o m ph ases ha help ou models o di e en ia e con ex .
Speci ically, we (i) eplace named en i ies such as company names, pe son, money alues o da es wi h p ede ined ags
using named en i y ecogni ion; (ii) pe o m lemma iza ion o educe wo ds o hei base o m; (iii) o m ph ases using
a da a-d i en app oach ha joins wo ds wi h signi ican co-occu ence (e.g., Mikolo , Su ske e , e al., 2013); and (i )
emo e s opwo ds (e.g., Li e al., 2021; Reichmann & Reichmann, 2022; Reichmann e al., 2022).10
2.5 Desc ip i e s a is ics
Table 3p o ides summa y s a is ics o ou sample. We epo ou main ou come a iable, common con ols in he
c ash isk li e a u e and ex ual cha ac e is ics o ou MD&A sample. The desc ip i e s a is ics sugges ha 23% o all
i m-yea s in ou sample expe ience a i m-speci ic s ock p ice c ash. Mo eo e , ou sample is cha ac e ized by g ow h
i ms as indica ed by a mean (median) ma ke - o-book a io (MTB) o 3.223 (2.229). The a e age i m-yea has inancial
le e age o 0.616 (0.475) a he sample mean (median).
We ind ha an MD&A con ains, on a e age, 1.2% nega i e wo ds (NEGATIVE). Only 0.4% o wo ds a e weak modal
wo ds (WEAK_MODAL), whe eas 1.4% a e associa ed wi h inancial unce ain y (UNCERTAIN), gene ally consis en
wi h Lough an and McDonald (2011). Mo eo e , he mean o he FOG index is 18.378 and adjus ing he FOG index
o inancial e ms ha a e ypically no conside ed complex by in es o s yields a subs an ially lowe a e age sco e o
13.284 (MODFOG). Collec i ely, ou sample esembles p e ious ones used in esea ch on c ash isk (e.g., E ug ul e al.,
2017; Kim e al., 2019).
3APPROACH TO PREDICTION
3.1 Sample pa i ioning
To ain and es ou models, we simula e a ealis ic o ecas ing scena io. Each model is ained on pas da a and hen
e alua ed on nex yea ’s hold-ou sample. Speci ically, ou app oach implemen s a olling sample spli ing scheme, in
which he aining and alida ion samples g adually shi o wa d in ime, bu he numbe o yea s in each sample is
held cons an (Chen e al., 2022).
8Choosing an uppe bound also depends on di e en aspec s such as he numbe o aining samples o he applied ex p ep ocessing s eps. When we
conside a olling spli o he es yea 2018 ha includes he i e p e ious yea s o model es ima ion, un abula ed esul s sugges ha an uppe bound o
40,000 wo ds conside s all wo ds ha appea mo e han i e imes in he sample o documen s. This sugges s ha choosing an uppe bound o he 40,000
mos equen ly occu ing wo ds seems easonable.
9To ain wo d2 ec, we chose a skip-g am app oach wi h a window size o i e, 20 i e a ions o e he co pus, a minimum wo d coun o i e and nega i e sam-
pling o accele a e aining (e.g., Li e al., 2021; Reichmann & Reichmann, 2022). To ensu e he alidi y o ou wo d2 ec embeddings, we pe o m desc ip i e
analyses in Appendix A1.
10 To be e illus a e ou ex p ep ocessing, conside he ollowing sen ence: “G oss ma gins dec eased as a esul o lowe sales in 2003.” A e ex
p ep ocessing he sen ence eads as ollows: “g oss_ma gin dec ease esul low sale -da e-.”
1102 KAYA ET AL.
TABLE 3 Desc ip i e s a is ics.
NMean S d. Q25 Median Q75
Ou come
CRASH +139,583 0.230 0.421 0.000 0.000 0.000
Common C ash Con ols
LOGMV 39,583 12.950 2.363 11.279 13.043 14.581
MTB 39,583 3.223 9.809 1.206 2.229 4.108
LEV 39,583 0.616 0.935 0.279 0.475 0.655
ROA 39,583 −0.004 0.318 −0.014 0.070 0.140
DTURN 39,583 0.002 0.109 −0.016 0.000 0.026
NCSKEW 39,583 0.116 0.927 −0.412 0.002 0.519
RET 39,583 −0.261 0.466 −0.258 −0.101 −0.040
SIGMA 39,583 0.057 0.045 0.028 0.045 0.072
MD&A Cha ac e is ics
LOGLENGTH 39,583 3.787 0.324 3.616 3.846 4.012
NEGATIVE 39,583 0.012 0.005 0.009 0.012 0.015
WEAK_MODAL 39,583 0.004 0.003 0.003 0.004 0.005
UNCERTAIN 39,583 0.014 0.005 0.011 0.014 0.017
FOG 39,583 18.376 1.711 17.327 18.326 19.338
MODFOG 39,583 13.284 1.573 12.317 13.224 14.160
LOGFILESIZE 39,583 0.052 0.031 0.028 0.048 0.070
No e: This able p esen s desc ip i e s a is ics o ou ou come a iable, common nume ical c ash con ols and MD&A cha ac-
e is ics. CRASH +1is an indica o a iable ha equals 1 i a i m-speci ic weekly e u n d ops 3.09 s anda d de ia ions below
i s yea ly mean in he pe iod +1 and 0 o he wise. All o he a iables a e de ined in Table 2.
Fo each yea in he es pe iod om 2001 o 2018 (e.g., 2001), he models a e ained in he i e p eceding yea s.
The i s ou yea s (e.g., 1996−1999) a e used o aining, and he i h yea is used o alida ion (e.g., 2000) o une
model hype pa ame e s (Chen e al., 2022). The p edic ion model is hen es ima ed on he alida ion and aining sam-
ple (e.g., 1996−2000) (Be omeu e al., 2021). Finally, we analyze he model pe o mance on he hold-ou - es sample
(e.g., 2011).
Al hough we acknowledge ha using a olling sample spli ing scheme is compu a ionally demanding, i has a leas
wo bene i s. Fi s , i s ic ly a oids empo al leakage as each model is ained on pas da a and hen es ed on nex
yea ’s hold-ou sample. Second, i allows he models o be con inuously ained on new da a.
3.2 Machine lea ning models
To p o ide a comp ehensi e analysis o he p edic abili y o s ock p ice c ashes, we conside a a ie y o di e en
models. Speci ically, we conside (i) a adi ional eg ession model, (ii) SVMs, (iii) models based on decision ees
and (i ) neu al ne wo ks. Table 4p o ides an o e iew o he models and hei espec i e da a inpu s. Mo eo e ,
o each olling spli , we sepa a ely une model hype pa ame e s using he alida ion se . Appendix A2 p o ides he
hype pa ame e space used o uning each o ou models.11
11 Fo mo e de ails on widely used machine lea ning me hods, we e e o Bochkay e al. (2023).
KAYA ET AL.1109
also conside ed impo an by he SGB(Num). In e es ingly, we ind ha inancial opaci y (OPAQUE), a s anda d con ol
a iable in he li e a u e (Hu on e al., 2009), only anks as he 22nd mos impo an nume ical p edic o , sugges ing
ela i ely low impo ance.
Tu ning o ex ual p edic o s, we conside he mos impo an e ms used by SGB(Tex ) o p edic s ock p ice
c ashes. We ind ha he wo d “g ow h” is he mos impo an p edic o in MD&A ex and ha i ms w i ing mo e
abou g ow h a e mo e likely o be c ash i ms as indica ed by he su ix (+). Mo eo e , e ms ela ed o in es men s
such as “acquisi ion,” “spending” and “p oduc ion” help ou model p edic s ock p ice c ashes. This no ion is consis en
wi h he iew ha high-g ow h i ms a e mo e likely o expe ience s ock p ice c ashes (e.g., Chen e al., 2001; Hu on
e al., 2009). Mo eo e , in line wi h p e ious esea ch sugges ing ha one is an impo an de e minan o c ash isk
(e.g., Fu e al., 2021; Reichmann, 2023), we ind ha SGB(Tex ) also cap u es e ms such as “downg ade,” “may ne e
succeed” and “un a o ably impac ” o dis inguish c ash om non-c ash obse a ions.
Collec i ely, ou esul s co obo a e ha ou models conside easonable signals o p edic s ock p ice c ashes.
Howe e , we cau ion he eade agains in e p e ing he esul s as indica i e o he causal in luence o p edic o s
(Chen e al., 2022). Ins ead, we aim o os e anspa ency and isualize unde lying da a inpu s ha d i e he p edic i e
pe o mance o ou main models.
4.3.3 Sensi i i y o se ial c ashes
In ou main es , we modi y he da a used o ain he model by se ing he alues o obse a ions o ze o i he same
i m is iden i ied as a c ash obse a ion in he ollowing yea ’s hold-ou sample. This design choice mi iga es conce ns
ha ou models simply p edic i ms ins ead o i m-yea s, he eby imp o ing he obus ness o ou main in e ences.
In his sec ion, we es he sensi i i y o ou models o his design choice.
Table 7shows he esul s o es ima ing all main models wi hou co ec ing o se ial c ashes in he aining da a. The
esul s in panel A sugges ha ailing o co ec o se ial c ashes in he aining da a in la es model pe o mance o
almos all models. This e ec is s onges o RF(Num)—an inc ease in AUC o 3.76%, almos doubling i s pe o mance
ela i e o a andom guess. By con as , he AUC o SGB(Num) only inc eases by 1.50%, sugges ing ha he model is
mo e likely o iden i y gene alizable pa e ns in inancial da a, compa ed o RF(Num).
In e es ingly, u ning o ex ual inpu s, he esul s in panel B sugges ha he model pe o mance o less lexible
models like LOGIT(Tex ) and SVM(Tex ) signi ican ly inc ease when ailing o co ec o se ial c ashes, sugges ing ha
hese models a e mo e likely o iden i y i m cha ac e is ics in ex a he han gene alizable language ha p edic s
s ock p ice c ashes. SGB(Tex ) is he only model wi h ex ual inpu s ha is no a ec ed by he co ec ion o se ial
c ashes. Ou esul s sugges ha ailing o co ec o se ial c ashes can subs an ially in la e he pe o mance o a -
ious model a chi ec u es. Ou esul s should cau ion u u e esea ch ha examines he p edic abili y o s ock p ice
c ashes.
5CONCLUSION
In his s udy, we use machine lea ning me hods d awn om he wide li e a u e in compu e science o p edic s ock
p ice c ashes. We es a ious models ha inco po a e nume ical and ex ual inpu s om 10-K disclosu es. We ind
ha a LOGIT model as a adi ional eg ession app oach se es as a s ong benchma k. We u he ind ha a SGB
model based on decision ees sys ema ically imp o es he p edic ion o one-yea -ahead s ock p ice c ashes. Ou
machine lea ning models a e mos aluable o ou -o -sample p edic ions using sui able combina ions o nume ical
and ex ual inpu s. We ind ha he mos powe ul ex ual p edic o s om MD&A sec ions a e such wo ds as “g ow h,”
“acquisi ion” and “spending.” The esul s should be o in e es o academics, p ac i ione s and in es o s who aim o
be e unde s and he p edic o s o s ock p ice c ashes. Fo ins ance, machine lea ning algo i hms can help in es o s
1110 KAYA ET AL.
TABLE 7 Sensi i i y o ecoding se ial c ashes.
Panel A: Nume ic models
Models AUC (main es ) AUC (w/o ecoding) Di . p- al
LOGIT(Num) 55.30% 55.87% +0.57% <0.01
SVM(Num) 52.90% 53.33% +0.43% <0.01
RF(Num) 53.42% 57.18% +3.76% <0.01
SGB(Num) 56.26% 57.76% +1.50% <0.01
NN(Num) 54.73% 54.97% +0.24% 0.4026
Panel B: Tex ual models
Models AUC (main es ) AUC (w/o ecoding) Di . p- al
LOGIT(Tex Cha ) 53.11% 53.62% +0.51% <0.01
LOGIT(Tex ) 52.84% 57.32% +4.48% <0.01
SVM(Tex ) 49.39% 55.61% +6.22% <0.01
RF(Tex ) 56.17% 57.93% +1.76% <0.01
SGB(Tex ) 56.18% 56.30% +0.12% 0.7087
CNN(Tex ) 54.77% 55.44% +0.67% 0.0144
No e: This able p esen s he esul s o p edic ing 1-yea -ahead s ock p ice c ashes wi hou ecoding se ial c ash obse a ions
in he aining se o model es ima ion. Panel A epo s he esul s o models using nume ical inpu s. Panel B epo s he
esul s o models using ex ual inpu s. “Num” deno es he 37 inancial a iables de ined in Table 2, panel A. “Tex Cha ” deno es
ex ual cha ac e is ics de ined in Table 2, panel B. “Tex ” deno es wo d2 ec embeddings o he CNN and TF-IDF weigh s o
he emaining ex models. The able epo s he AUC. Di e ences be ween AUC sco es a e es ed o s a is ical signi icance
using he DeLong es .
o posi ion hei po olios agains u u e s ock p ice c ashes and hus help hem make be e -in o med in es men
decisions.
While ou s udy p o ides ea ly e idence on he use o machine lea ning me hods o p edic ing s ock p ice c ashes,
mo e e idence is needed. S ock p ice c ashes a e p e alen in in e na ional inancial ma ke s, ye he inc emen al p e-
dic i e powe o o he (manda o y) i m disclosu es and da a sou ces, such as sus ainabili y epo s, social media da a
and consume p oduc e iews o u u e s ock p ice c ashes emain la gely unexplo ed (e.g., Al Guindy e al., 2024;
El-Haj e al., 2020;Jin,2023).
ACKNOWLEDGMENTS
This pape is pa ially based on Chap e 2 o Do on Reichmann’s disse a ion comple ed a Ruh Uni e si y Bochum,
and Chap e 2 was p e iously i led “P edic ing Fi m-Speci ic S ock P ice C ashes.” We hank pa icipan s a he
Leipzig Banking and Finance Wo kshop, HVB Semina in Pade bo n, FAACT and i u Semina in Bochum, Ge man Aca-
demic Associa ion o Business Resea ch Con e ence, Uni e si y o E langen-Nü nbe g, wo anonymous e iewe s
o he EAA Con e ence, And ew S a k (senio edi o ) and an anonymous e iewe a he Jou nal o Business Finance
& Accoun ing and Jonas Ewe z, Pe oula Glach siou, Tho s en Knaue , Cha lo e Knick ehm, Johannes K iebel,
Rou en Mölle , Ma in Nienhaus, S ephan Paul, Be nha d Pellens, Ma hias Pels e , And eas P ings en, Fleming
Schmid -Skipiol, And é Uhde and G ego Weiss o help ul commen s on ea lie e sions o his pape .
Open access unding enabled and o ganized by P ojek DEAL.
DATA AVAILABILITY STATEMENT
Da a a e a ailable om he sou ces as ci ed in he ex .
KAYA ET AL.1111
ORCID
Do on Reichmann h ps://o cid.o g/0000-0002-2196-1746
REFERENCES
Al Guindy, M., Naugh on, J. P., & Rio dan, R. (2024). The e olu ion o co po a e wi e usage. Jou nal o Business Finance &
Accoun ing,51(3-4), 819–845. h ps://doi.o g/10.1111/jb a.12758
An, Z., Chen, Z., Li, D., & Xing, L. (2018). Indi idualism and s ock p ice c ash isk. Jou nal o In e na ional Business S udies,49(9),
1208–1236. h ps://doi.o g/10.1057/s41267-018-0150-z
Baesens, B., Van Ges el, T., Viaene, S., S epano a, M., Suykens, J., & Van hienen, J. (2003). Benchma king s a e-o - he-a clas-
si ica ion algo i hms o c edi sco ing. Jou nal o he Ope a ional Resea ch Socie y,54(6), 124–136. h ps://doi.o g/10.1016/
j.ejo .2015.05.030
Bali, T. G., Beckmeye , H., Mö ke, M., & Weige , F. (2023). Op ion e u n p edic abili y wi h machine lea ning and big da a. The
Re iew o Financial S udies,36(9), 3548–3602. h ps://doi.o g/10.1093/ s/hhad017
Be omeu, J., Cheynel, E., Floyd, E., & Wenqiang, P. (2021). Using machine lea ning o de ec miss a emen s. Re iew o
Accoun ing S udies,26, 468–519. h ps://doi.o g/10.1007/s11142-020-09563-8
Bianchi, D., Büchne , M., & Tamoni, A. (2020). Bond isk p emiums wi h machine lea ning. The Re iew o Financial S udies,34(2),
1046–1089. h ps://doi.o g/10.1093/ s/hhaa062
Bochkay, K., B own, S. V., Leone, A. J., & Tucke , J. W. (2023). Tex ual analysis in accoun ing: Wha ’s nex ? Con empo a y
Accoun ing Resea ch,40(2), 765–805. h ps://doi.o g/10.1111/1911-3846.12825
B eiman, L. (2001). Random o es s. Machine Lea ning,45, 5–32. h ps://doi.o g/10.1023/A:1010933404324
B own, S. V., & Tucke , J. W. (2011). La ge-sample e idence on i ms’ yea -o e -yea MD&A modi ica ions. Jou nal o Accoun ing
Resea ch,49(2), 309–346. h ps://doi.o g/10.1111/j.1475-679X.2010.00396.x
Bu a u, F., Chen, Q., Cla k, B., Das, S., Lo, A. W., & Siddique, A. (2016). Risk and isk managemen in he c edi ca d indus y.
Jou nal o Banking & Finance,72, 218–239. h ps://doi.o g/10.1016/j.jbank in.2016.07.015
Chen, J., Hong, H., & S ein, J. C. (2001). Fo ecas ing c ashes: T ading olume, pas e u ns, and condi ional skewness in s ock
p ices. Jou nal o Financial Economics,61(3), 345–381. h ps://doi.o g/10.1016/S0304-405X(01)00066-6
Chen, X., Cho, Y. H. (Tony), Dou, Y., & Le , B. (2022). P edic ing u u e ea nings changes using machine lea ning and de ailed
inancial da a. Jou nal o Accoun ing Resea ch,60(2), 467–515. h ps://doi.o g/10.1111/1475-679X.12429
Cheng, C., Jones, S., & Mose , W. J. (2018). Abno mal ading beha io o speci ic ypes o sha eholde s be o e US i m
bank up cy and i s implica ions o i m bank up cy p edic ion. Jou nal o Business Finance & Accoun ing,45(9-10),
1100–1138. h ps://doi.o g/10.1111/jb a.12338
Dechow, P. M., Ge, W., La son, C. R., & Sloan, R. G. (2011). P edic ing ma e ial accoun ing miss a emen s. Con empo a y
Accoun ing Resea ch,28(1), 17–82. h ps://doi.o g/10.1111/j.1911-3846.2010.01041.x
DeLong, E. R., DeLong, D. M., & Cla ke-Pea son, D. L. (1988). Compa ing he a eas unde wo o mo e co ela ed ecei e
ope a ing cha ac e is ic cu es: A nonpa ame ic app oach. Biome ics,44(3), 837–845.
Du, Z., Huang, A. G., We me s, R., & Wu, W. (2022). Language and domain speci ici y: A Chinese inancial sen imen dic iona y.
Re iew o Finance,26(3), 673–719. h ps://doi.o g/10.1093/ o / ab036
Eaglesham, J. (2013, May 27). SEC e ocuses on accoun ing aud. Wall S ee Jou nal.h p://online.wsj.com/a icle/
SB10001424127887324125504578509241215284044.h ml
El-Haj, M., Al es, P., Rayson, P., Walke , M., & Young, S. (2020). Re ie ing, classi ying and analysing na a i e commen a y in
uns uc u ed (glossy) annual epo s published as PDF iles. Accoun ing and Business Resea ch,50(1), 6–34. h ps://doi.o g/
10.1080/00014788.2019.1609346
El-Haj, M., Rayson, P., Walke , M., Young, S., & Simaki, V. (2019). In sea ch o meaning: Lessons, esou ces and nex s eps o
compu a ional analysis o inancial discou se. Jou nal o Business Finance & Accoun ing,46(3-4), 265–306. h ps://doi.o g/
10.1111/jb a.12378
E ug ul, M., Lei, J., Qiu, J., & Wan, C. (2017). Annual epo eadabili y, one ambigui y, and he cos o bo owing. The Jou nal
o Financial and Quan i a i e Analysis,52(2), 811–836. h ps://doi.o g/10.1017/S0022109017000187
Fawce , T. (2006). An in oduc ion o ROC analysis. Pa e n Recogni ion Le e s,27(8), 861–874. h ps://doi.o g/10.1016/j.
pa ec.2005.10.010
F iedman, J. H. (2002). S ochas ic g adien boos ing. Compu a ional S a is ics & Da a Analysis,38(4), 367–378. h ps://doi.o g/
10.1016/S0167-9473(01)00065-2
Fu, X., Wu, X., & Zhang, Z. (2021). The in o ma ion ole o ea nings con e ence call one: E idence om s ock p ice c ash isk.
Jou nal o Business E hics,173(3), 643–660. h ps://doi.o g/10.1007/s10551-019-04326-1
Gu, S., Kelly, B., & Xiu, D. (2020). Empi ical asse p icing ia machinelea ning. The Re iew o Financial S udies,33(5), 2223–2273.
h ps://doi.o g/10.1093/ s/hhaa009
Hin on, G. E., Osinde o, S., & Teh, Y. W. (2006). A as lea ning algo i hm o deep belie ne s. Neu al Compu a ion,18(7), 1527–
1554.
1112 KAYA ET AL.
Hong, H. A., Kim, J. B., & Welke , M. (2017). Di e gence o cash low and o ing igh s, opaci y, and s ock p ice c ash isk:
In e na ional e idence. Jou nal o Accoun ing Resea ch,55(5), 1167–1212. h ps://doi.o g/10.1111/1475-679X.12185
Hu on, A. P., Ma cus, A. J., & Teh anian, H. (2009). Opaque inancial epo s, R2, and c ash isk. Jou nal o Financial Economics,
94(1), 67–86. h ps://doi.o g/10.1016/j.j ineco.2008.10.003
Jin, L., & Mye s, S. C. (2006). R2a ound he wo ld: New heo y and new es s. Jou nal o Financial Economics,79(2), 257–292.
h ps://doi.o g/10.1016/j.j ineco.2004.11.003
Jin, S. (2023). Mo e han wo ds: Can one o consume p oduc e iews help p edic i ms’ undamen als? Jou nal o Business
Finance & Accoun ing,50(9-10), 1910–1942. h ps://doi.o g/10.1111/jb a.12680
Jones, S., & Henshe , D. A. (2004). P edic ing i m inancial dis ess: A mixed logi model. The Accoun ing Re iew,79(4), 1011–
1038. h ps://doi.o g/10.2308/acc .2004.79.4.1011
Jones, S., & Henshe , D. A. (2007). Modelling co po a e ailu e: A mul inomial nes ed logi analysis o uno de ed ou comes.
The B i ish Accoun ing Re iew,39(1), 89–107. h ps://doi.o g/10.1016/j.ba .2006.12.003
Jones, S., Johns one, D., & Wilson, R. (2015). An empi ical e alua ion o he pe o mance o bina y classi ie s in he p edic ion
o c edi a ings changes. Jou nal o Banking & Finance,56, 72–85. h ps://doi.o g/10.1016/j.jbank in.2015.02.006
Jones, S., Johns one, D., & Wilson, R. (2017). P edic ing co po a e bank up cy: An e alua ion o al e na i e s a is ical
amewo ks. Jou nal o Business Finance & Accoun ing,44(1-2), 3–34. h ps://doi.o g/10.1111/jb a.12218
Jones, S., Mose , W. J., & Wieland, M. M. (2023). Machine lea ning and he p edic ion o changes in p o i abili y. Con empo a y
Accoun ing Resea ch,40(4), 2643–2672. h ps://doi.o g/10.1111/1911-3846.12888
Kaya, D., & Seebeck, A. (2019). The dissemina ion o i m in o ma ion ia company egis e websi es: Coun y-le el empi ical
e idence. Jou nal o Accoun ing & O ganiza ional Change,15(3), 382–429. h ps://doi.o g/10.1108/JAOC-03-2018-0023
Kim, C. (F ancis), Wang, K., & Zhang, L. (2019). Readabili y o 10-K epo s and s ock p ice c ash isk. Con empo a y Accoun ing
Resea ch,36(2), 1184–1216. h ps://doi.o g/10.1111/1911-3846.12452
Kim, J.-B., Li, Y., & Zhang, L. (2011a). CFOs e sus CEOs: Equi y incen i es and c ashes. Jou nal o Financial Economics,101(3),
713–730. h ps://doi.o g/10.1016/j.j ineco.2011.03.013
Kim, J.-B., Li, Y., & Zhang, L. (2011b). Co po a e ax a oidance and s ock p ice c ash isk: Fi m-le el analysis. Jou nal o Financial
Economics,100(3), 639–662. h ps://doi.o g/10.1016/j.j ineco.2010.07.007
Kunche a, L. I., Bezdek, J. C., & Duin, R. P. W. (2001). Decision empla es o mul iple classi ie usion: An expe imen al
compa ison. Pa e n Recogni ion,34(2), 299–314. h ps://doi.o g/10.1016/S0031-3203(99)00223-X
LeCun, Y., Bengio, Y., & Hin on, G. (2015). Deep lea ning. Na u e,521(7553), 436–444.
LeCun, Y., Bose , B., Denke , J. S., Hende son, D., Howa d, R. E., Hubba d, W., & Jackel, L. D. (1989). Backp opaga ion applied o
handw i en zip code ecogni ion. Neu al Compu a ion,1(4), 541–551.
Lewis, C., & Young, S. (2019). Fad o u u e? Au oma ed analysis o inancial ex and i s implica ions o co po a e epo ing.
Accoun ing and Business Resea ch,49(5), 587–615. h ps://doi.o g/10.1080/00014788.2019.1611730
Li, K., Mai, F., Shen, R., & Yan, X. (2021). Measu ing co po a e cul u e using machine lea ning. The Re iew o Financial S udies,
34(7), 3265–3315. h ps://doi.o g/10.1093/ s/hhaa079
Li, L., Jamieson, K., DeSal o, G., Ros amizadeh, A., & Talwalka , A. (2017). Hype band: A no el bandi -based app oach o
hype pa ame e op imiza ion. Jou nal o Machine Lea ning Resea ch,18(1), 6765–6816.
Lough an, T., & McDonald, B. (2011). When is a liabili y no a liabili y? Tex ual analysis, dic iona ies, and 10-Ks. The Jou nal o
Finance,66(1), 35–65. h ps://doi.o g/10.1111/j.1540-6261.2010.01625.x
Lough an, T., & McDonald, B. (2014). Measu ing eadabili y in inancial disclosu es. The Jou nal o Finance,69(4), 1643–1671.
h ps://doi.o g/10.1111/jo i.12162
Mai, F., Tian, S., Lee, C., & Ma, L. (2019). Deep lea ning models o bank up cy p edic ion using ex ual disclosu es. Eu opean
Jou nal o Ope a ional Resea ch,274(2), 743–758. h ps://doi.o g/10.1016/j.ejo .2018.10.024
Mikolo , T., Chen, K., Co ado, G., & Dean, J. (2013). E icien es ima ion o wo d ep esen a ions in ec o space.h ps://doi.o g/
10.48550/a xi .1301.3781
Mikolo , T., Su ske e , I., Chen, K., Co ado, G., & Dean, J. (2013). Dis ibu ed ep esen a ions o wo ds and ph ases and hei
composi ionali y.h ps://doi.o g/10.48550/a xi .1310.4546
Pea , M., & Jones, S. (2012). Using neu al ne s o combine in o ma ion se s in co po a e bank up cy p edic ion. In elligen
Sys ems in Accoun ing,Finance and Managemen ,19(2), 90–101. h ps://doi.o g/10.1002/isa .334
Rapach, D. E., S auss, J. K., & Zhou, G. (2010). Ou -o -sample equi y p emium p edic ion: combina ion o ecas s and links o
he eal economy. Re iew o Financial S udies,23(2), 821–862. h ps://doi.o g/10.1093/ s/hhp063
Reichmann, D., & Reichmann, M. (2022). P edic ing i m-speci ic s ock p ice c ashes. In D. Reichmann, Quali a i e disclosu es
and capi al ma ke consequences—A ex ual analysis app oach. [Doc o al disse a ion, Ruh Uni e si y Bochum].
Reichmann, D. (2023). Tone managemen and s ock p ice c ash isk. Jou nal o Accoun ing and Public Policy,42(6), 107155.
h ps://doi.o g/10.1016/j.jaccpubpol.2023.107155
Reichmann, D., Mölle , R., & He el, T. (2022). No hing bu good in en ions: The sea ch o equi y and s ock p ice c ash isk.
Jou nal o Business Economics,92(9), 1455–1489. h ps://doi.o g/10.1007/s11573-022-01085-w
KAYA ET AL.1113
Schmid , P. S., on A x, U., Sch imp , A., Wagne , A. F., & Ziegle , A. (2019). Common isk ac o s in in e na ional s ock ma ke s.
Financial Ma ke s and Po olio Managemen ,33, 213–241. h ps://doi.o g/10.1007/s11408-019-00334-3
Shumway, T. (2001). Fo ecas ing bank up cy mo e accu a ely: A simple haza d model. The Jou nal o Business,74(1), 101–124.
h ps://doi.o g/10.1086/209665
Vapnik, V. N. (1998). S a is ical lea ning heo y. Wiley-In e science.
Wu, K., & Lai, S. (2020). In angible in ensi y and s ock p ice c ash isk. Jou nal o Co po a e Finance,64, 101682. h ps://doi.o g/
10.1016/j.jco p in.2020.101682
How o ci e his a icle: Kaya, D., Reichmann, D., & Reichmann, M. (2025). Ou -o -sample p edic abili y o
i m-speci ic s ock p ice c ashes: A machine lea ning app oach. Jou nal o Business Finance & Accoun ing,52,
1095–1115. h ps://doi.o g/10.1111/jb a.12831
1114 KAYA ET AL.
APPENDIX
A.1 Wo d2 ec alida ion
In his sec ion, we ensu e he alidi y o ou wo d2 ec embeddings (Reichmann & Reichmann, 2022). Speci ically, we
p o ide desc ip i e e idence on how di e en wo d g oups occupy di e en loca ions in he ec o space. Figu e A1
shows wo d2 ec embeddings in a h ee-dimensional sca e plo , which sugges s ha closely ela ed wo ds such as
“p o i ,” “ma gin” and “sales” occupy close loca ions. Fu he , wo d g oups ela ed o go e nance, such as “independen
audi o ” and “in e nal con ol,” end o be close in he ec o space. We in e ha wo d2 ec embeddings iden i y
easonable seman ic simila i ies (Reichmann & Reichmann, 2022).
FIGURE A1 Visualiza ion o wo d2 ec embeddings. This igu e p esen s a isualiza ion o ou wo d2 ec
embeddings. We employ -dis ibu ed s ochas ic neighbo embedding echniques o isualize he 300-dimensional
wo d2 ec embeddings in a h ee-dimensional sca e plo . The do s ep esen he educed ec o ep esen a ions o
wo ds and ph ases con ained in ou sample o MD&A o 10-K ilings (n=39,583).
A.2 Model hype pa ame e s
This sec ion p esen s he hype pa ame e s es ed o ain ou machine lea ning models. Fo he LOGIT and SVM, we
ollow he speci ica ions o Mai e al. (2019) wi hou pe o ming u he op imiza ion. Fo he decision ees, RF and
SGB, we choose simila hype pa ame e s as in Chen e al. (2022) and use g id sea ch o iden i y op imal hype pa ame-
e s. Simila o Gu e al. (2020), we es di e en laye dep hs and he numbe o neu ons pe laye o bo h he NN and
CNN and implemen ea ly-s opping o p e en o e i ing du ing aining. In addi ion, we also es a ange o ac i a ion
unc ions and hype pa ame e s ha a e speci ic o he CNN. Because aining he neu al ne wo ks is compu a ionally
KAYA ET AL.1115
expensi e, we use he Hype band app oach o une hype pa ame e s ha aims o speed up andom sea ch h ough
adap i e esou ce alloca ion and ea ly-s opping (e.g., Li e al., 2017).
Model Hype pa ame e s Op imiza ion
Logis ic eg ession (LOGIT) L1 egula iza ion –
Suppo ec o machines
(SVM)
Radial basis unc ion ke nel –
Random o es (RF) # ees: 500, 600, 700, ...,2,000 G idsea ch
Max ea u es: 110, 111, 112, ...,120
Min. # o obs. in a lea : 10
Bagging: 0.5
S ochas ic g adien boos ing
(SGB)
# ees: 500, 600, 700, ...,2,000 G id sea ch
Lea ning a e: 0.005, 0.01, 0.05
Max.dep h:1,2,3,4
Min. # o obs. in a lea : 10
Bagging: 0.5
Neu al ne wo k (NN) # hidden laye : 1,2,3,4,5 Hype band
# neu ons in hidden laye s: 50, 100, 200, 300 400, 500
Ac i a ion unc ion (e e y node excep inal): elu, anh,
sigmoid
Regula ize : loa in [0, 0.0001]
Con olu ional neu al ne wo k
(CNN)
# hidden laye : 1,2,3,4,5 Hype band
# neu ons in hidden laye s: 50, 100, 200, 300 400, 500
# il e : 200, 250, 300, ...,500
ke nelsize:2,4,6,8,...,20
Embedding dimension (wo d2 ec): 50, 100, 200, 300
Ac i a ion unc ion: elu, anh, sigmoid
Regula ize : loa in [0, 0.0001]
NN-CNN (conca laye ) # neu ons in hidden laye : 50, 100, 200, 300 400, 500 Hype band
No e: This able p esen s he hype pa ame e s and op imiza ion echniques used o ain ou machine lea ning models.
