1145
Online ISSN: 3006-2047
P in ISSN: 3006-2039
The Impac o A i icial In elligence on Financial Fo ecas ing
Accu acy in Co po a e Finance
Muhammad Kam an
BS G adua e om he Depa men o Accoun ing and Finance Koha Uni e si y o
Science and Technology, KPK Pakis an
Email: [email p o ec ed]
A nan Shahid
BS G adua e in Bachelo o Business adminis a ion in Finance om Kushal khan
Kha ak Uni e si y Ka ak, KPK Pakis an
Email: [email p o ec ed]
Abs ac
This s udy explo es how a i icial in elligence imp o es inancial o ecas ing accu acy
wi hin co po a e inance and compa es he pe o mance o adi ional s a is ical
models wi h mode n machine lea ning and deep lea ning echniques. Using a olling-
o igin e alua ion om 2014 o 2024, he analysis examines sho - and long-ho izon
o ecas s o e enue, ope a ing cash low, and ea nings. The esul s show ha AI
models, pa icula ly XGBoos , LSTM, and ensemble app oaches, consis en ly deli e
lowe o ecas ing e o s and emain s able du ing shi s in economic condi ions.
T adi ional models pe o m easonably well in sho windows bu lose accu acy when
ma ke ola ili y inc eases o when he o ecas ing ho izon ex ends. S a is ical
signi icance es s con i m ha he gains achie ed by AI models a e meaning ul and
no due o chance. The indings indica e ha i ms ha in eg a e AI-d i en o ecas ing
in o hei planning p ocesses can s eng hen budge ing, educe unce ain y, and
suppo mo e dependable long- e m decisions.
Keywo ds: A i icial In elligence, Fo ecas ing Accu acy, Co po a e Finance,
Machine Lea ning, Financial Modelling
In oduc ion
A i icial in elligence (AI) is apidly ans o ming co po a e inance by enhancing he
p ecision o inancial o ecas ing. Mode n o ecas ing me hods, adi ionally d i en
by s a is ical models o expe judgmen , o en s uggle o cap u e he complex,
nonlinea ela ionships and eal- ime dynamics p esen in inancial ma ke s. By
le e aging machine lea ning and deep lea ning echniques, AI can p ocess massi e
da ase s, de ec sub le pa e ns, and adap o shi ing economic condi ions, he eby
imp o ing o ecas accu acy and educing e o a es. Fo co po a ions, mo e accu a e
o ecas s mean be e capi al alloca ion, op imised isk managemen , and mo e
in o med s a egic planning.
1146
Online ISSN: 3006-2047
P in ISSN: 3006-2039
Empi ical esea ch suppo s his ans o ma ion. In a s udy o i ms lis ed on he
Pakis an S ock Exchange, AI-d i en o ecas ing ools signi ican ly imp o ed
p edic ion accu acy by enabling he de ec ion o nonlinea ends and eal- ime ma ke
signals. (Khan e al., 2025) Mo eo e , AI models a e shown o ou pe o m
con en ional s a is ical me hods by educing o ecas ing e o s, hanks o hei abili y
o con inuously lea n and ecalib a e. (Das e al., 2025) Resea ch has also ound ha
he inco po a ion o AI in o inancial analy ics no only bols e s o ecas accu acy bu
also s eng hens isk managemen capabili ies, esul ing in mo e esilien co po a e
decision-making. (Mbonigaba & Mish a, 2025) A sys ema ic li e a u e e iew u he
indica es ha AI’s ole in eal- ime o ecas ing helps i ms p oac i ely espond o
ma ke ola ili y and mi iga e downside isks. (U ama & Hidaya , 2025) A he same
ime, he adop ion o AI in co po a e inance is accele a ing: neu al ne wo ks, suppo
ec o machines, and ensemble models a e now widely used, and hey deli e gains in
p edic i e powe a beyond adi ional econome ic echniques. (Sadiq, Adeel &
Luqman, 2025)
Howe e , he shi is no wi hou challenges. Ba ie s such as da a quali y,
in e p e abili y o AI models, in as uc u al cons ain s, and egula o y unce ain y
emain signi ican hu dles o widesp ead adop ion. (Asi e al., 2025) These issues
poin o a nuanced ela ionship be ween AI adop ion and o ecas ing accu acy: while
he po en ial bene i s a e subs an ial, success ul implemen a ion equi es s a egic
in es men in echnological capabili ies and o ganisa ional eadiness.
Resea ch Objec i es
To examine how AI-based models imp o e he accu acy o inancial o ecas ing
compa ed o adi ional o ecas ing me hods in co po a e inance.
To iden i y he key AI echniques used in co po a e inancial o ecas ing and e alua e
hei e ec i eness
To explo e he challenges companies ace when adop ing AI o inancial o ecas ing
and how hese challenges a ec o ecas ing accu acy.
Resea ch Ques ions
How do AI-based inancial o ecas ing models imp o e o ecas ing accu acy
compa ed o adi ional me hods in co po a e inance?
Which AI echniques a e commonly used in co po a e inancial o ecas ing, and how
e ec i e a e hey in imp o ing accu acy?
Wha challenges do companies ace when implemen ing AI o inancial o ecas ing,
and how do hese challenges in luence he accu acy o o ecas s?
Signi icance o he S udy
This s udy is impo an because i helps explain how a i icial in elligence can
s eng hen inancial o ecas ing in co po a e inance a a ime when i ms a e dealing
wi h as -mo ing ma ke s and la ge olumes o inancial da a. Mo e accu a e
o ecas s suppo be e in es men decisions, imp o e isk managemen , and help
companies plan wi h g ea e con idence. The esea ch also highligh s which AI
1147
Online ISSN: 3006-2047
P in ISSN: 3006-2039
echniques o e he mos alue and iden i ies a eas whe e o ganisa ions may
encoun e challenges when implemen ing hem. Unde s anding hese poin s can guide
inancial manage s, policy make s, and co po a e leade s as hey in es in new
echnology and wo k o imp o e he eliabili y o hei inancial o ecas s.
Li e a u e Re iew
A i icial in elligence (AI) and machine lea ning (ML) ha e become cen al o
con empo a y app oaches o inancial o ecas ing in co po a e inance. T adi ional
o ecas ing me hods, classical ime-se ies models such as ARIMA o exponen ial
smoo hing and judgmen al o ecas ing emain use ul in many con ex s, bu hey
encoun e limi s when da a a e high-dimensional, nonlinea , o when aluable signals
a e la en in uns uc u ed sou ces (e.g., ex , images, al e na i e da a). AI me hods
anging om ee-based ensembles o deep neu al ne wo ks o e he abili y o exploi
ich ea u e se s, model nonlinea in e ac ions, and upda e dynamically as new da a
a i e. A g owing empi ical and e iew li e a u e indica es ha , when p ope ly
applied, AI me hods can ma e ially imp o e o ecas accu acy o a a ie y o
co po a e o ecas ing asks, al hough hose gains depend on p oblem aming, da a
quali y, model e alua ion igou , and o ganisa ional eadiness (Vancsu a, 2025;
Hyndman & A hanasopoulos, 2018).
Me hods and model amilies. The li e a u e ypically g oups o ecas ing app oaches
in o h ee amilies: classical s a is ical/econome ic models; “shallow” machine
lea ning models ( o example, ee ensembles and suppo ec o machines); and deep
lea ning models ( o example, ecu en and con olu ional a chi ec u es). Classical
models such as ARIMA, exponen ial smoo hing, and s a e-space me hods a e well
unde s ood, easy o in e p e , and o en pe o m well when he se ies ha e s able
pa e ns and small numbe s o p edic o s. Machine lea ning models, andom o es s
and g adien boos ed ees (no ably XGBoos ) handle la ge p edic o se s, in e ac ions,
and nonlinea i y and a e widely used in p ac ice because o hei s ong ou -o -sample
pe o mance and compu a ional scalabili y (Chen & Gues in, 2016). Deep lea ning
models, pa icula ly ecu en neu al ne wo ks (RNNs) and long sho - e m memo y
(LSTM) ne wo ks, excel a cap u ing complex empo al dependencies when la ge
olumes o da a a e a ailable (Hoch ei e & Schmidhube , 1997; Fawaz e al., 2018).
Hyb id and ensemble s a egies ha combine s a is ical and ML app oaches a e
equen ly ecommended, since no single me hod consis en ly domina es ac oss
o ecas ing p oblems (Fawaz e al., 2018; Chen & Gues in, 2016).
E idence on accu acy imp o emen s. Sys ema ic e iews and empi ical s udies
gene ally ind measu able accu acy gains o AI me hods ac oss many o ecas ing
domains (Vancsu a, 2025). Fo co po a e inance asks, cash low, e enue, ea nings,
c edi isk, and liquidi y o ecas s, ML and deep lea ning app oaches o en educe
common e o me ics (e.g., RMSE, MAE, MAPE) ela i e o naï e o pu ely
s a is ical baselines when da ase s a e su icien ly la ge and p edic o s a e in o ma i e
(Gu, Kelly, & Xiu, 2020). Howe e , he magni ude and obus ness o imp o emen s
a y. Se e al me a-analyses and ca e ul compa a i e s udies epo ha imp o emen s
sh ink when models a e es ed on uly ou -o -sample pe iods ha include s uc u al
1148
Online ISSN: 3006-2047
P in ISSN: 3006-2039
b eaks, egime shi s, o c isis episodes; gains a e also smalle when esea che s ail
o con ol o look-ahead bias o da a leakage (Hyndman & A hanasopoulos, 2018;
Vancsu a, 2025). Thus, epo ed AI bene i s a e condi ional on e alua ion igou and
da a p ope ies.
Why AI can help in p ac ice. Two echnical s eng hs explain much o AI’s ad an age.
Fi s , mode n ML me hods can inges mul imodal da a nume ical accoun ing ime
se ies, high- equency ansac ions, ex ual disclosu es (ea nings calls, MD&A), news
and sen imen indica o s, and al e na i e signals such as web sea ch ends, and lea n
complex c oss-modal ela ionships (Gu e al., 2020). NLP and ans o me -based
me hods can con e managemen commen a y, analys epo s, and egula o y ilings
in o quan i a i e ea u es ha imp o e p edic i e powe o ea nings and e en
o ecas ing. Second, ensemble and hyb id pipelines educe a iance and bias h ough
model a e aging o cascading (e.g., eeding esiduals om s a is ical models in o ML
models), which o en enhances obus ness (Chen & Gues in, 2016; Fawaz e al.,
2018).
E alua ion p ac ices and pi alls. The li e a u e emphasises igo ous e alua ion as a
key de e minan o c edible esul s. P ope ime-se ies c oss- alida ion ( olling
windows), a oidance o look-ahead bias, and epo ing mul iple me ics a e essen ial.
Many compa a i e s udies ha ini ially epo ed la ge gains o DL o ML me hods
la e e ised expec a ions a e in oducing s ic ou -o -pe iod es s and obus ness
checks (Hyndman & A hanasopoulos, 2018). Re iews call o s anda dised
benchma k da ase s, anspa en epo ing o hype pa ame e s and uning p ocedu es,
and open eplica ion code so ha claimed accu acy gains a e e i iable (Vancsu a,
2025).
Domain-speci ic applica ions in co po a e inance. AI applica ions in co po a e
inance include budge and cash- low o ecas ing, sho - e m e enue o ecas ing,
ea nings p edic ion, c edi isk sco ing, and o ecas ing ma ke esponses o co po a e
communica ions. ML me hods a e especially use ul whe e ansac ional o ope a ional
da a a e ich (e.g., e ail sales ac oss SKUs, paymen pipeline da a), enabling g anula
o ecas s ha eed olling budge ing and wo king-capi al managemen (Gu e al.,
2020). Empi ical s udies ac oss egions, No h Ame ica, Eu ope, and Sou h Asia,
epo simila pa e ns: AI me hods ou pe o m adi ional app oaches unde
a ou able da a condi ions bu a e sensi i e o da a equency, se ies ola ili y, and
ea u e enginee ing quali y (Vancsu a, 2025).
Role o ex ual and policy signals. Tex ual analysis has eme ged as an impo an
augmen a ion in o ecas ing. Models ha pa se cen al bank s a emen s, managemen
discussion & analysis (MD&A), and ea nings call ansc ip s ex ac sen imen and
opic ea u es ha add p edic i e in o ma ion o policy-sensi i e ou comes and i m
pe o mance. S udies ind ha combining nume ic accoun ing p edic o s wi h ex ual
indica o s yields consis en imp o emen s in o ecas s o ea nings su p ises and
ce ain sho -ho izon ou comes (Gu e al., 2020; Fawaz e al., 2018).
Compa a i e s udies and ensembles. Compa a i e wo k shows clea con ex s in which
ensembles o hyb ids ou pe o m single app oaches. Fo many co po a e o ecas ing
asks, ensembles ha combine ARIMA- ype baselines wi h boos ed ees o neu al
1149
Online ISSN: 3006-2047
P in ISSN: 3006-2039
ne s p oduce lowe e o and mo e s able p edic ions ac oss egimes (Chen &
Gues in, 2016; Fawaz e al., 2018). These indings encou age p ac i ione s o adop
model-s acking and assembling a he han sea ching o a single “bes ” algo i hm.
In e p e abili y, us , and go e nance. Fo ecas accu acy alone does no gua an ee
adop ion. In e p e abili y is c ucial in co po a e en i onmen s whe e o ecas s d i e
esou ce alloca ion and isk managemen . The li e a u e on explainable AI (XAI)
supplies ools, local su oga e models (LIME), SHAP alue decomposi ions, and
global ea u e-impo ance analyses o explain and audi complex models (Ribei o e
al., 2016; Lundbe g & Lee, 2017; Doshi-Velez & Kim, 2017). These me hods help
manage s and audi o s unde s and model d i e s, de ec biases, and sa is y
go e nance equi emen s. Many case s udies indica e ha he p esence o
in e p e able explana ions is a decisi e ac o in manage ial accep ance e en when a
“black-box” model deli e s sligh ly highe accu acy (Doshi-Velez & Kim, 2017;
Lundbe g & Lee, 2017).
Da a quali y, ea u e enginee ing, and domain expe ise. Mul iple e iews s ess ha
gains om AI depend hea ily on he ups eam wo k: da a cleaning, handling missing
alues, accoun ing o co po a e epo ing i egula i ies, and designing domain-
ele an ea u es (Hyndman & A hanasopoulos, 2018; Vancsu a, 2025). Al e na i e
da a o en b ings signal bu also noise; he cos o acquisi ion and he challenges in
alida ion a e non i ial. Domain expe ise in accoun ing and co po a e ope a ions is,
he e o e, essen ial o ans o m aw eco ds in o p edic i e ea u es ha co ec ly
cap u e business cycles and accoun ing p ac ices.
O e i ing, sample size, and obus ness. Deep models a e powe ul bu p one o
o e i ing when se ies a e sho o when aining igno es empo al dependence. The
li e a u e ecommends olling c oss- alida ion, egula isa ion, ea ly s opping, and
assembling as de ences. Me a-s udies no e ha pe o mance ad an ages epo ed in-
sample o en dissipa e in genuine ou -o -pe iod es s, pa icula ly when s uc u al
b eaks occu (Fawaz e al., 2018; Vancsu a, 2025). Acco dingly, c edible claims
abou AI imp o ing o ecas ing accu acy mus be suppo ed by obus ness ac oss
mul iple pe iods and s ess scena ios.
Compu a ional cos and deploymen challenges. T aining s a e-o - he-a neu al
ne wo ks o la ge ensembles can be compu a ionally and inancially expensi e,
equi ing specialised ha dwa e and enginee ing suppo . Deploymen issues,
con inuous moni o ing, e aining, model e sioning, and in eg a ion wi h ERP and
FP&A sys ems a e signi ican o ganisa ional hu dles (Chen & Gues in, 2016; Gu e
al., 2020). The li e a u e s esses ha i ms o en unde es ima e he o al cos o
owne ship o AI o ecas ing sys ems and ha success ul deploymen equi es c oss-
unc ional eams ha combine da a scien is s, inance p o essionals, and IT/De Ops
capabili ies.
Regula ion, audi abili y, and e hics. Regula o y and audi conce ns a e inc easingly
salien , pa icula ly when o ecas s eed c edi decisions o a ec in es o
communica ions. Regula o s p e e ep oducible pipelines, documen ed assump ions,
and ai ness es ing. The li e a u e sugges s ha hyb id models o explainable
1150
Online ISSN: 3006-2047
P in ISSN: 3006-2039
al e na i es can s ike a balance be ween accu acy and egula o y accep abili y
(Doshi-Velez & Kim, 2017; Lundbe g & Lee, 2017).
Manage ial implica ions and bes p ac ices. Fo inance p o essionals, he li e a u e
o e s p agma ic guidance: in es in obus da a go e nance and in as uc u e be o e
adop ing complex models; p io i ise o ecas ing p oblems wi h equen obse a ions
and ich p edic o s; combine accu acy gains wi h in e p e abili y ools and human
o e sigh ; and adop inc emen al pilo s ha emphasise ope a ional in eg a ion a he
han isola ed model p oo -o -concep s (Hyndman & A hanasopoulos, 2018; Vancsu a,
2025). Case s udies show ha i ms ha embed AI o ecas s wi hin decision
wo k lows a he han ea ing hem as s andalone ou pu s ealise he mos ope a ional
alue.
Open ques ions and u u e di ec ions. The li e a u e poin s o se e al p omising
esea ch di ec ions: benchma king co po a and eplica ion s udies o co po a e
o ecas ing asks; ans e lea ning and p e aining applied o inancial ime se ies;
imp o ed in e p e abili y me hods ailo ed o inance p ac i ione s; and longi udinal
s udies ha quan i y he long- un impac o AI deploymen s on o ecas ing accu acy,
manage ial decisions, and i m ou comes (Gu e al., 2020; Vancsu a, 2025). Resea ch
ha connec s o ecas ing imp o emen s o measu able business bene i s (e.g., educed
wo king-capi al cos s, lowe o ecas bias, imp o ed ea nings guidance) would be
pa icula ly use ul o p ac i ione s.
O e all, he li e a u e p esen s a cau iously op imis ic assessmen : AI me hods o en
enhance o ecas ing accu acy in co po a e inance when applied in sui able con ex s
and igo ously e alua ed. The ealised bene i s depend on da a a ailabili y and quali y,
he choice o model and ensemble s a egy, in e p e abili y and go e nance conce ns,
and he i m’s ope a ional capabili ies o deploy and moni o models. Fu u e wo k
should emphasise anspa en benchma king, domain- ocused in e p e abili y, and
e idence linking o ecas imp o emen s o manage ial and inancial ou comes.
Resea ch Me hodology
Resea ch Design
This s udy ollows a quan i a i e, compa a i e esea ch design. The goal is o examine
how a i icial in elligence models pe o m agains adi ional o ecas ing app oaches
in p edic ing key co po a e inance indica o s such as e enue, ope a ing cash low,
and ea nings pe sha e. The design ocuses on measu ing o ecas ing accu acy,
compa ing model pe o mance ac oss mul iple ho izons, and es ing whe he he
imp o emen s obse ed in AI models a e s a is ically signi ican . A olling-o igin
e alua ion amewo k was selec ed because i mi o s eal o ecas ing en i onmen s
whe e i ms con inuously upda e models as new inancial da a becomes a ailable.
This app oach also cap u es he e ec o changing economic condi ions on model
pe o mance.
Da a Sou ces
The analysis uses publicly a ailable inancial da a om co po a e qua e ly epo s
co e ing he pe iod om 2014 o 2024. Th ee p ima y indica o s we e selec ed:
1151
Online ISSN: 3006-2047
P in ISSN: 3006-2039
e enue, ope a ing cash low, and ea nings pe sha e. These indica o s we e chosen
because hey a e cen al o co po a e planning, budge ing, and in es o
communica ion. Addi ional ma ke a iables, such as s ock p ices, we e collec ed o
cap u e ex e nal signals associa ed wi h ea nings cycles. All da a we e d awn om
eliable inancial da abases, including company ilings and ma ke da a eposi o ies.
The selec ion ensu es ha he da ase e lec s eal co po a e epo ing pa e ns and
includes bo h s able and ola ile economic pe iods.
Sampling Technique and Time F ame
A pu posi e sampling s a egy was applied o selec i ms wi h comple e da a o e he
10 yea s. Consis en qua e ly obse a ions we e necessa y o model aining, olling-
o igin es ing, and c oss-ho izon e alua ion. The ime ame co e s expansion cycles,
down u ns, and eco e y phases, p o iding a sui able en i onmen o es ing how
models beha e unde changing condi ions. This pe iod also includes yea s in which
AI adop ion in co po a e analy ics accele a ed, making he compa ison be ween AI-
based and adi ional models mo e ele an .
Da a P ep ocessing
Be o e model de elopmen , he da ase was cleaned and p epa ed o a oid biases and
imp o e eliabili y. Missing alues we e impu ed using o wa d o backwa ds illing
o sho gaps and model-based impu a ion o longe gaps. All a iables we e
inspec ed o s a iona i y, and log ans o ma ions we e applied whe e necessa y.
Time indices we e aligned o p e en look-ahead bias. Seasonal pa e ns we e
cap u ed h ough qua e and mon h indica o s, and lagged ea u es we e added o
imp o e p edic i e s eng h. Fo models ha use ex e nal a iables, addi ional
p edic o s such as ma ke ola ili y and basic sen imen sco es om ea nings-call
ansc ip s we e included. These s eps ollow s anda d p ac ices in o ecas ing and
ensu e ha each model ecei es a consis en and ai se o inpu s.
Model De elopmen
Fou amilies o o ecas ing models we e de eloped:
T adi ional Models
Naï e o ecas (ca y- o wa d)
ARIMA/SARIMA wi h au oma ic o de selec ion
These models se e as baselines o e alua ing he alue added by AI.
Machine Lea ning Models
Random Fo es
G adien Boos ed T ees (XGBoos )
These models handle nonlinea ela ionships and in e ac ions ha adi ional me hods
miss.
1152
Online ISSN: 3006-2047
P in ISSN: 3006-2039
Deep Lea ning Models
Long Sho -Te m Memo y (LSTM)
Hyb id LSTM–ARIMA model
These app oaches cap u e long- e m dependencies and nonlinea s uc u es in
inancial ime se ies.
Ensemble Model
A simple a e age and s acked ensemble we e included because ensembles o en
inc ease obus ness by combining he s eng hs o di e en models. Each model was
ained using he same aining windows, ea u es, and o ecas ing ho izons o ensu e
a ai compa ison.
Fo ecas E alua ion F amewo k
Fo ecas accu acy was assessed using mul iple e o me ics:
Mean Absolu e E o (MAE)
Roo Mean Squa ed E o (RMSE)
Mean Absolu e Pe cen age E o (MAPE)
Using mul iple me ics helps cap u e di e en aspec s o accu acy, such as sensi i i y
o ou lie s, e o magni ude, and pe cen age-based in e p e a ion.
The e alua ion included h ee o ecas ing ho izons:
One qua e ahead
Th ee qua e s ahead
Fou qua e s ahead
This design allowed he s udy o compa e how models pe o m in sho - e m
budge ing e sus longe - e m s a egic planning.
C oss-Valida ion and Ou -o -Sample Tes ing
A olling-o igin e alua ion me hod was used o simula e eal-wo ld o ecas ing.
Models we e i s ained on da a om 2014–2018 and used o gene a e 2019
o ecas s. The aining window was hen olled o wa d by one yea , and new
o ecas s we e p oduced. This p ocess con inued un il 2024. This me hod ensu es ha
models a e always es ed on unseen da a and ha pe o mance is e alua ed unde
changing ma ke condi ions.
S a is ical Signi icance Tes ing
To de e mine whe he imp o emen s in accu acy we e meaning ul a he han
coinciden al, he Diebold–Ma iano (DM) es was applied. This es compa es o ecas
e o s be ween wo compe ing models while adjus ing o se ial co ela ion in ime-
se ies da a. Each AI model was compa ed wi h ARIMA as he benchma k. Addi ional
pai ed - es s o Wilcoxon signed- ank es s we e used whe e e o dis ibu ions
equi ed hem. Models we e conside ed supe io i he p- alue ell below he .05
h eshold.
1153
Online ISSN: 3006-2047
P in ISSN: 3006-2039
E hical Conside a ions
The s udy used publicly a ailable inancial da a, which a oids conce ns ela ed o
con iden iali y o pe sonal p i acy. No sensi i e o p op ie a y in o ma ion was
accessed. The analysis was conduc ed anspa en ly, and all s eps a e eplicable. Ca e
was aken o epo esul s objec i ely and a oid o e s a ing he pe o mance o any
model.
Limi a ions
Al hough he me hodology is igo ous, ce ain limi a ions emain. The s udy ocuses
on qua e ly da a, which may hide sho - e m luc ua ions ha daily o weekly da a
could e eal. AI models equi e la ge da ase s, and esul s migh a y o smalle
i ms wi h limi ed epo ing his o ies. Finally, he analysis e alua es only selec ed AI
models; di e en a chi ec u es o addi ional ea u es could u he change accu acy
ou comes.
Da a Analysis
This sec ion p esen s he da a analysis ca ied ou o examine how a i icial
in elligence models compa e wi h adi ional o ecas ing app oaches in co po a e
inance. The goal is o e alua e accu acy, eliabili y, and consis ency ac oss di e en
o ecas ho izons and ma ke condi ions. The analysis e lec s es ablished p ac ices in
inancial o ecas ing esea ch and ollows he amewo k ou lined in he me hodology
chap e . The ocus is on h ee se s o models: adi ional s a is ical models, machine-
lea ning models, and deep-lea ning models. Thei accu acy is assessed h ough
s anda d e o me ics and o mal s a is ical es s. The chap e also explains how hese
models beha e unde changing condi ions, how hey espond o inpu ea u es, and
how manage s can in e p e hei ou pu s.
E en hough i ms use di e en in e nal da a sys ems, his analysis uses a s uc u ed
da ase con aining qua e ly e enue, ope a ing cash low, and ea nings pe sha e o a
sample o companies o e en yea s. This mi o s wha appea s in empi ical
o ecas ing esea ch and allows o a balanced e alua ion o con en ional and AI-
d i en app oaches.
O e iew o he Da a
The da ase includes qua e ly obse a ions om 2014 o 2024. Each se ies con ains
ac ual alues and co esponding o ecas o igins o h ee key indica o s: e enue,
ope a ing cash low, and ea nings pe sha e. These indica o s we e selec ed because
hey a e widely used in planning, capi al budge ing, and ea nings guidance. The
da ase also includes a small se o ex e nal ea u es: ma ke e u ns, in e es - a e
changes, c ude-oil p ice shi s, and sen imen ex ac ed om manage ial discussion
sec ions in qua e ly ilings. These ea u es we e added o es how AI models espond
o iche in o ma ion compa ed o adi ional me hods, which usually ocus only on
pas obse a ions.
Be o e analysis, all se ies we e checked o consis ency. Missing alues we e a e,
a ec ing only a ew qua e s o a hand ul o i ms. These gaps we e handled by
1160
Online ISSN: 3006-2047
P in ISSN: 3006-2039
Das, S., Tulsyan, A., e al. (2025). AI-powe ed p edic i e analy ics in inancial
o ecas ing. In e na ional Jou nal o Inno a i e Science and Ad anced
Enginee ing (IJISAE).
Doshi-Velez, F., & Kim, B. (2017). Towa ds a igo ous science o in e p e able
machine lea ning. a Xi . h ps://a xi .o g/abs/1702.08608
Fawaz, H. I., Fo es ie , G., Webe , J., Idoumgha , L., & Mulle , P.-A. (2018). Deep
lea ning o ime se ies classi ica ion: A e iew. a Xi .
h ps://a xi .o g/abs/1809.04356
Gu, S., Kelly, B., & Xiu, D. (2020). Empi ical asse p icing ia machine lea ning. The
Re iew o Financial S udies, 33(5), 2223–2273.
h ps://doi.o g/10.1093/ s/hhaa009
Hoch ei e , S., & Schmidhube , J. (1997). Long sho - e m memo y. Neu al
Compu a ion, 9(8), 1735–1780. h ps://doi.o g/10.1162/neco.1997.9.8.1735
Hyndman, R. J., & A hanasopoulos, G. (2018). Fo ecas ing: P inciples and p ac ice
(2nd ed.). OTex s. h ps://o ex s.com/ pp2/
Khan, A., Ali, R., & Ahmed, Z. (2025). The ole o a i icial in elligence in enhancing
inancial o ecas ing accu acy. Asian Jou nal o Economics, Accoun ing and
Finance (AJEAF). h ps://ajea .com
Lundbe g, S. M., & Lee, S.-I. (2017). A uni ied app oach o in e p e ing model
p edic ions. In Ad ances in Neu al In o ma ion P ocessing Sys ems, 30 (pp.
4765–4774).
Mbonigaba, J., & Mish a, T. (2025). AI-d i en inancial analy ics: Enhancing o ecas
accu acy and isk managemen . Nepal Jou nal o Managemen (NepJol).
Ribei o, M. T., Singh, S., & Gues in, C. (2016). “Why should I us you?”:
Explaining he p edic ions o any classi ie . In P oceedings o he 22nd ACM
SIGKDD In e na ional Con e ence on Knowledge Disco e y and Da a Mining
(pp. 1135–1144).
Sadiq, M., Adeel, M., & Luqman, A. (2025). Le e aging AI in inancial o ecas ing:
Cu en ends and u u e di ec ions. Resea ch Jou nal o Managemen and
Social Sciences (RJMSS). h ps:// jmssjou nal.com
U ama, Y., & Hidaya , W. (2025). The ole o a i icial in elligence in inancial
o ecas ing: A sys ema ic li e a u e e iew. Jou nal o Public Policy and
Business Re iew (PPIPBR). h ps://jou nal.ppipb .com
Vancsu a, L. (2025). Na iga ing AI-d i en inancial o ecas ing: A sys ema ic e iew
o cu en s a us and c i ical esea ch gaps. Jou nal o Financial Technology &
Fo ecas ing, 7(3), 36–52.
