Uni e sidade do Minho
Escola de Engenha ia
Diogo Gonçal es G açoei o
The Impo ance o Wo k o ce Planning
and Tu no e P edic ion in Re ail:
Le e aging People Analy ics o S a egic
Decision Making
Ab il de 2025
Uni e sidade do Minho
Escola de Engenha ia
Diogo Gonçal es G açoei o
The Impo ance o Wo k o ce Planning
and Tu no e P edic ion in Re ail:
Le e aging People Analy ics o S a egic
Decision Making
Disse ação de Mes ado
Mes ado em Engenha ia e Ges ão Indus ial
T abalho e e uado sob a o ien ação da
P o esso a Ma ia do Samei o Ca alho
P o esso João Nuno Gonçal es
Ab il de 2025
ii
DIREITOS DE AUTOR E CONDIÇÕES DE UTILIZAÇÃO DO TRABALHO POR TERCEIROS
Es e é um abalho académico que pode se u ilizado po e cei os desde que espei adas as
eg as e boas p á icas in e nacionalmen e acei es, no que conce ne aos di ei os de au o e di ei os
conexos.
Assim, o p esen e abalho pode se u ilizado nos e mos p e is os na licença abaixo indicada.
Caso o u ilizado necessi e de pe missão pa a pode aze um uso do abalho em condições não
p e is as no licenciamen o indicado, de e á con ac a o au o , a a és do Reposi ó iUM da Uni e sidade
do Minho.
Licença concedida aos u ilizado es des e abalho
A ibuição
CC BY
h ps://c ea i ecommons.o g/licenses/by/4.0/
iii
ACKNOWLEDGMENTS
As I each he culmina ion o my academic pa h, I am o e whelmed wi h a p o ound sense o
accomplishmen and g a i ude. Comple ing his hesis means he end o a ema kable chap e in my li e
and ep esen s he ul illmen o a long-held aspi a ion. This ans o ma i e jou ney has en iched me wi h
in aluable expe iences.
A i s wo d o he Uni e si y o Minho, whe e I pu sued my deg ee, o p o iding me wi h a
comp ehensi e educa ion and os e ing an en i onmen o academic excellence. The knowledge and skills
I acqui ed du ing my s udies ha e been pi o al in shaping my ca ee pa h. I would like o exp ess my
hea el hanks o P o esso Ma ia Ca alho and P o esso João Gonçal es, my supe iso s a he
Uni e si y. Thei unwa e ing suppo , aluable guidance and insigh ul eedback ha e signi ican ly
in luenced he di ec ion and quali y o my esea ch.
Fo emos , I would like o exp ess my deepes g a i ude o LTPlabs, he company whe e I had he
p i ilege o comple e my in e nship. Thei unwa e ing suppo and guidance h oughou my in e nship
we e ins umen al in shaping my skills and knowledge in he ield. I am immensely hank ul o all he
LTPee s o hei wa m welcome, collabo a ion and sha ed expe ise, which g ea ly con ibu ed o my
p o essional g ow h.
i
STATEMENT OF INTEGRITY
I he eby decla e ha ing conduc ed his academic wo k wi h in eg i y. I con i m ha I ha e no
used plagia ism o any o m o undue use o in o ma ion o alsi ica ion o esul s along he p ocess
leading o i s elabo a ion.
I u he decla e ha I ha e ully acknowledged he Code o E hical Conduc o he Uni e si y o
Minho.
A Impo ância do Planeamen o da Fo ça de T abalho e da P e isão de Ro a i idade no Re alho: Ala anca People Analy ics
pa a a Tomada de Decisão Es a égica
RESUMO
O p esen e abalho p opõe uma abo dagem ino ado a pa a o planeamen o es a égico de
ecu sos humanos no se o do e alho, a a és do desen ol imen o de modelos p edi i os de
u no e
com base em écnicas de People Analy ics. A ele ada axa de u no e nes e se o , aliada à necessidade
de uma ges ão mais e icien e da o ça de abalho, jus i icou a c iação de duas soluções dis in as, ambas
baseadas no algo i mo XGBoos : um modelo de p e isão a um ano e ou o a cinco anos.
O modelo de um ano em como obje i o es ima a p obabilidade de saída olun á ia de cada
colabo ado com base nas suas ca ac e ís icas indi iduais. Já o modelo de cinco anos é desen ol ido a
pa i de dados ag egados, conside ando combinações de ca ac e ís icas como unção, localização e ipo
de loja, bem como pe is- ipo de colabo ado es com a ibu os semelhan es. Es a es u u a pe mi e
ealiza p ojeções es a égicas de longo p azo em con ex os de maio ince eza. Pa a assegu a a
aplicabilidade p á ica das p e isões, o modelo u iliza apenas a iá eis e ca ego ias cujo alo é conhecido
ou es imá el à pa ida.
Ambos os modelos demons a am capacidades p edi i as ele an es, des acando-se a
impo ância de a iá eis como emune ação ixa, senio idade e idade. No en an o, e i ica am-se
en iesamen os sis emá icos em alguns segmen os — especialmen e nos g upos com meno senio idade
— e uma u ilização limi ada das a aliações de pe o mance e po encial, a o es que o am analisados
c i icamen e. As p e isões desen ol idas nes a disse ação des inam-se a alimen a um modelo de
o imização já exis en e, e o çando o seu po encial pa a apoia decisões es a égicas de ecu sos
humanos.
Com base nos esul ados ob idos, o es udo ap esen a ambém um conjun o de p opos as pa a
in es igação e desen ol imen o u u o, que incluem a e isão da o ma como os pe is- ipo de
colabo ado es são de inidos, a inco po ação da mobilidade in e na e o e o ço da in e p e abilidade local
do modelo. Es e p oje o ep esen a um con ibu o p á ico e ele an e pa a a adoção de abo dagens
da a-
d i en
na ges ão de alen o, ap oximando os ecu sos humanos da es a égia o ganizacional.
PALAVRAS-CHAVE
People Analy ics; Re alho; Tu no e ; Machine Lea ning; Planeamen o Es a égico de Recu sos Humanos
i
ii
The Impo ance o Wo k o ce Planning and Tu no e P edic ion in Re ail: Le e aging People Analy ics o S a egic Decision
Making
ABSTRACT
This disse a ion p oposes an inno a i e app oach o s a egic wo k o ce planning in he e ail
sec o h ough he de elopmen o p edic i e u no e models based on People Analy ics echniques. The
high u no e a e in his sec o , combined wi h he need o mo e e icien wo k o ce managemen ,
jus i ied he c ea ion o wo dis inc solu ions, bo h using he XGBoos algo i hm: a one-yea p edic ion
model and a i e-yea o ecas model.
The one-yea model aims o es ima e he p obabili y o olun a y employee exi based on
indi idual cha ac e is ics. The i e-yea model, on he o he hand, is buil om agg ega ed da a,
conside ing combina ions o ea u es such as ole, loca ion and s o e ype, as well as employee- ype
p o iles wi h simila a ibu es. This s uc u e enables long- e m s a egic p ojec ions in con ex s o highe
unce ain y. To ensu e p ac ical applicabili y, he model elies only on a iables and ca ego ies whose
u u e alues a e known o can be eliably es ima ed in ad ance.
Bo h models demons a ed ele an p edic i e capabili ies, wi h ixed compensa ion, senio i y
and age s anding ou as key d i e s. Howe e , sys ema ic biases we e iden i ied in some segmen s —
pa icula ly among ea ly-ca ee g oups — and pe o mance and po en ial e alua ions showed limi ed
in luence, which we e c i ically assessed. These p edic ions a e designed o se e as inpu o an exis ing
wo k o ce op imiza ion model, suppo ing mo e in o med hi ing and e en ion planning o e ime.
Based on he esul s, he s udy also p esen s a se o p oposals o u u e esea ch and
de elopmen , including a e ision o how employee g oups a e de ined, he inco po a ion o in e nal
mobili y and enhanced local in e p e abili y o he model. This p ojec p o ides a p ac ical and ele an
con ibu ion o he adop ion o da a-d i en app oaches in alen managemen , b idging human esou ces
and o ganiza ional s a egy.
KEYWORDS
People Analy ics; Re ail; Tu no e ; Machine Lea ning; S a egic Wo k o ce Planning
iii
x
ACRONYMS
AI – A i icial In elligence.
AUC – A ea Unde he Cu e.
AUC-ROC – A ea Unde he Recei e Ope a ing Cha ac e is ic Cu e.
FTE – Full-Time Equi alen .
GBM – G adien Boos ing Machine.
GBT – G adien Boos ing T ee.
GLM – Gene alized Linea Model.
HR – Human Resou ces.
HRIS – Human Resou ce In o ma ion Sys ems.
HRM – Human Resou ce Managemen .
KNN – K-Nea es Neighbo s.
LDA – Linea Disc iminan Analysis.
LR – Logis ic Reg ession.
MAE – Mean Absolu e E o .
ML – Machine Lea ning.
NN – Neu al Ne wo k.
RF – Random Fo es .
RNN – Recu en Neu al Ne wo k.
RQ – Resea ch Ques ion.
SHAP - SHapley Addi i e exPlana ions.
SMOTE – Syn he ic Mino i y O e sampling Technique.
SVM – Suppo Vec o Machine.
VBM – Value Based Managemen .
XGBoos – Ex eme G adien Boos ing.
1
1. INTRODUCTION
The p esen disse a ion is wi hin he ealm o he Mas e in Indus ial Enginee ing and
Managemen a he Uni e si y o Minho. The hesis was de eloped du ing an in e nship a LTPLabs, an
analy ical managemen consul ancy company, as pa o a p ojec wi h one o i s clien s. The in en o
his chap e is o gi e a gene al o e iew o he disse a ion’s subjec and ex en , p oblems ha a e
handled, p ojec ’s objec i es and s uc u e o he en i e documen .
1.1 Mo i a ion
Employee u no e emains a pe sis en challenge in he e ail sec o , a ec ing ope a ional e iciency,
cus ome se ice quali y and o e all business pe o mance. High u no e a es lead o inc eased
ec ui men and aining cos s while also dis up ing s o e ope a ions, pa icula ly in on line oles whe e
expe ience and con inui y a e c i ical o main aining se ice s anda ds (Olubiyi e al., 2019). Despi e
ad ancemen s in Human Resou ces (HR) echnologies — such as wo k o ce analy ics ools and alen
managemen pla o ms — wo k o ce planning in many o ganiza ions con inues o ely hea ily on empi ical
knowledge a he han da a-d i en decision-making, limi ing he abili y o an icipa e wo k o ce needs
e ec i ely (Majam & Ja bandhan, 2022).
T adi ional wo k o ce planning is o en eac i e, adjus ing s a ing le els only a e u no e has
occu ed. In con as , People Analy ics enables a p oac i e app oach, le e aging his o ical wo k o ce da a
— such as enu e, ole ansi ions and demog aphic indica o s — o de elop p edic i e models capable o
o ecas ing u no e pa e ns (Polze , 2022). By in eg a ing p edic i e analy ics in o wo k o ce planning,
companies can imp o e long- e m s a ing s a egies, op imize hi ing p ocesses and ensu e business
con inui y (Majam & Ja bandhan, 2022).
Beyond wo k o ce planning, o ganiza ional cul u e also in luences u no e dynamics. Olubiyi e
al. (2019) emphasize ha Pe son-O ganiza ion Fi signi ican ly impac s job sa is ac ion and employee
commi men , which in u n a ec s e en ion. Employees who align wi h company alues and pe cei e a
s ong sense o belonging a e mo e likely o emain, pa icula ly in cus ome - acing en i onmen s whe e
engagemen di ec ly in luences se ice quali y. Addi ionally, hei s udy highligh s ha u no e is d i en
no only by inancial incen i es bu also by ac o s such as leade ship quali y, ca ee de elopmen
oppo uni ies and wo kplace cul u e.
2
While exis ing esea ch explo es a ious ac o s in luencing u no e , ewe s udies ha e
examined how p edic i e analy ics can be sys ema ically applied o wo k o ce planning. Machine lea ning
models a e inc easingly being adop ed in HR decision-making, o e ing g ea e accu acy in iden i ying
wo k o ce ends and in o ming s a egic planning (Heidemann e al., 2024). Howe e , he e hical
implica ions o A i icial In elligence (AI)-d i en HR analy ics mus also be conside ed o ensu e ha
p edic i e models a e implemen ed in a esponsible and anspa en manne ha suppo s bo h
o ganiza ional and employee in e es s (Gie mindl e al., 2022).
This disse a ion add esses his gap by de eloping a p edic i e u no e model ailo ed o he
e ail sec o . Ra he han ocusing on indi idual e en ion s a egies, he s udy aims o suppo s a egic
wo k o ce planning by o ecas ing employee depa u es using machine lea ning echniques. The
p edic i e model p o ides annual u no e es ima es, allowing o a mo e da a-d i en app oach in s a ing
decisions. By iden i ying key wo k o ce a iables and analyzing u no e ends, his esea ch o e s a
s uc u ed me hodology o in eg a ing p edic i e analy ics in o HR decision-making. The indings will help
HR p o essionals an icipa e wo k o ce gaps, op imize hi ing needs and imp o e cos e iciency in
wo k o ce managemen .
1.2 P ojec Backg ound
This disse a ion was de eloped du ing a consul ing p ojec o a majo Po uguese company
ope a ing in he e ail sec o , speci ically in he consume goods segmen . The company is a key playe
in bo h he na ional and in e na ional ma ke , managing a la ge po olio o s o es and o e ing a di e se
ange o p oduc s, including g oce ies, household i ems and pe sonal ca e p oduc s. Wi h an ex ensi e
s o e ne wo k and a high ansac ion olume, he company se es millions o cus ome s annually ac oss
di e en loca ions, equi ing ca e ul wo k o ce managemen o main ain ope a ional e iciency.
Gi en he dynamic na u e o consume demand and he complexi y o managing a la ge wo k o ce,
ensu ing s a egic wo k o ce planning is a key challenge. Fo his s udy, he ocus was placed on h ee
p ima y s o e o ma s ha cons i u e he co e o he company's ope a ions:
• La ge hype ma ke s, cha ac e ized by high ansac ion olumes and equi ing a la ge, s able
wo k o ce;
• Medium-sized supe ma ke s, which balance p oduc a ie y and e iciency, necessi a ing
p ecise wo k o ce planning;
• Neighbo hood con enience s o es, whe e agili y and wo k o ce lexibili y a e c i ical o
ope a ions.
3
Despi e s uc u ed wo k o ce planning e o s, employee u no e emains a majo challenge,
a ec ing wo k o ce s abili y, se ice quali y and ope a ional cos s. Cu en ly, u no e p edic ions a e
based on empi ical me hods, elying on his o ical wo k o ce a ia ions a he han p edic i e analy ics.
This app oach lacks analy ical dep h and limi s he company's abili y o an icipa e wo k o ce gaps
p oac i ely. Add essing his gap in u no e o ecas ing is essen ial o enhancing wo k o ce planning and
op imizing esou ce alloca ion, ensu ing ha s o es emain well-s a ed o mee business demands.
1.3 P ojec Objec i es and Expec ed Resul s
The p ima y objec i e o his disse a ion is o de elop a u no e p edic ion model using
ad anced machine lea ning echniques o suppo s a egic wo k o ce planning. Gi en he company's
eliance on his o ical wo k o ce a ia ions o es ima e u no e , his model seeks o in oduce a da a-
d i en app oach ha inco po a es bo h in insic (e.g., employee a ibu es) and ex insic (e.g., ma ke
condi ions) ac o s.
The model will gene a e yea -by-yea u no e p edic ions o e a i e-yea pe iod, p o iding
g anula insigh s in o employee depa u e p obabili ies. Unlike he exis ing empi ical app oach, which
p ima ily uses Full-Time Equi alen (FTE)-based es ima ions, his model will allow he company o
p oac i ely add ess wo k o ce gaps, pa icula ly in key ope a ional unc ions.
F om he second yea onwa ds, he model does no aim o p edic indi idual exi s bu a he
o ecas s he expec ed numbe o olun a y depa u es agg ega ed by employee p o iles, ole clus e ,
s o e ype and egion — suppo ing g oup-le el wo k o ce planning ins ead o employee-le el o ecas ing.
The expec ed ou come o his disse a ion is he de elopmen o a high-p ecision u no e
p edic ion model, capable o iden i ying pa e ns in employee a i ion and p o iding ac ionable insigh s
o suppo HR planning and decision-making. This p edic i e capabili y will no only enable HR and
managemen eams o de elop a ge ed e en ion s a egies and adjus ec ui men plans, bu will also
se e as a key inpu o an al eady exis ing wo k o ce op imiza ion model de eloped by he company o
p ojec eam. Th ough his in eg a ion, he p edic ed u no e a es will suppo mo e e icien wo k o ce
s uc u ing, helping o balance hi ing, in e nal mobili y and p omo ion decisions ac oss di e en employee
segmen s and geog aphical loca ions.
In addi ion o model de elopmen , his disse a ion also aims o e lec on how p edic i e ools
and People Analy ics can add ess he challenges o managing olun a y u no e in high- a iabili y
en i onmen s, such as e ail and how such ools can suppo mo e e ec i e, da a-d i en wo k o ce
s a egies.
4
To guide he in es iga ion and s uc u e he na a i e ac oss he di e en chap e s, he ollowing
esea ch ques ions (RQs) a e p oposed:
• RQ1: Wha a e he main indi idual and con ex ual ac o s ha in luence olun a y employee
u no e in he e ail sec o ?
• RQ2: How can p edic i e analy ics be used o o ecas olun a y u no e and suppo
s a egic wo k o ce planning in high- u no e en i onmen s?
1.4 Thesis Ou line
The p esen disse a ion is o ganized in o six main chap e s, each con ibu ing o a s uc u ed
unde s anding o he u no e p edic ion p oblem in he e ail sec o and he de elopmen o a da a-d i en
solu ion. Chap e 1 in oduces he hesis by p esen ing he mo i a ion, p ojec backg ound, objec i es
and expec ed esul s, as well as ou lining he documen s uc u e. Chap e 2 p o ides a li e a u e e iew
ha con ex ualizes he issue o employee u no e and highligh s he ele ance o wo k o ce planning and
p edic i e analy ics. I also explo es adi ional and mode n app oaches o u no e p edic ion, including
machine lea ning me hods, wi h special emphasis on he XGBoos algo i hm. Chap e 3 desc ibes he
p oblem add essed in his p ojec , beginning wi h an analysis o he cu en decision-making con ex ,
iden i ying he limi a ions o he exis ing app oach and p o iding a de ailed o e iew o he a ailable da a
and explo a o y analysis. Chap e 4 ou lines he me hodological app oach adop ed o de elop he
p edic i e models. I includes de ails on da a p epa a ion, cleaning, p ocessing and s uc u ing, ollowed
by he deploymen o wo models: a one-yea indi idual-le el model and a i e-yea agg ega e-le el model.
Chap e 5 p esen s he esul s ob ained om each model. I e alua es hei p edic i e pe o mance and
in e p e s he mos impo an a iables d i ing u no e , based on hei ele ance in he XGBoos
amewo k. Chap e 6 deli e s a c i ical analysis o he models’ s eng hs and limi a ions, ollowed by a
se o sho - e m and medium/long- e m ecommenda ions o u u e esea ch and de elopmen . The
disse a ion concludes wi h e e ences and an appendix ha includes addi ional analyses and da a used
o suppo he wo k.
5
2. LITERATURE REVIEW
This chap e p o ides a s uc u ed li e a u e e iew o suppo he de elopmen o a p edic i e model
o employee u no e in he e ail sec o . I begins by cla i ying he concep o u no e and i s
o ganiza ional implica ions, ollowed by a e iew o adi ional app oaches and he key in insic and
ex insic ac o s ha in luence olun a y u no e . The chap e hen explo es how wo k o ce planning can
bene i om p edic i e analy ics, highligh ing bo h i s s a egic ele ance and p ac ical challenges. Finally,
a ious machine lea ning echniques a e analyzed, wi h a special ocus on decision ee models such as
XGBoos , o e alua e hei sui abili y o u no e p edic ion in high- a iabili y en i onmen s. This e iew
lays he heo e ical ounda ion o he me hodology adop ed in his disse a ion.
2.1 In oduc ion o Employee Tu no e
Employee u no e e e s o he a e a which employees lea e an o ganiza ion and mus be
eplaced. I is a c i ical wo k o ce me ic ha signi ican ly impac s o ganiza ional s abili y, inancial
pe o mance and ope a ional e iciency (Olubiyi e al., 2019). Tu no e can occu o a ious easons,
anging om job dissa is ac ion and ca ee p og ession o o ganiza ional es uc u ing and layo s, making
i a key conce n in Human Resou ce Managemen (HRM) (Al-Su aihi e al., 2021).
The e a e wo p ima y ypes o u no e : olun a y and in olun a y. Volun a y u no e occu s
when employees choose o lea e hei jobs, o en in sea ch o be e ca ee oppo uni ies, highe sala ies,
imp o ed wo k-li e balance o g ea e job sa is ac ion. On he o he hand, in olun a y u no e happens
when employees a e e mina ed due o ac o s such as layo s, unde pe o mance o s uc u al changes
wi hin he company (Al-Su aihi e al., 2021). Addi ionally, u no e can be classi ied as unc ional o
dys unc ional. Func ional u no e e e s o he depa u e o unde pe o ming employees, which can
bene i he o ganiza ion by imp o ing eam e iciency, whe eas dys unc ional u no e in ol es he loss o
high-pe o ming employees, leading o po en ial skill sho ages and dis up ions in business ope a ions
(Wallace & Gaylo , 2012). Ano he impo an dis inc ion is be ween a oidable and una oidable u no e .
A oidable u no e esul s om ac o s ha o ganiza ions can con ol, such as poo managemen , lack
o ca ee g ow h o inadequa e compensa ion, while una oidable u no e is d i en by ex e nal ac o s,
such as e i emen , pe sonal eloca ion o heal h issues (Ba ick & Zimme man, 2005).
Taking u no e in o accoun is essen ial because o i s inancial, ope a ional and cul u al
implica ions. Financially, eplacing an employee is cos ly, wi h expenses ela ed o ec ui men ,
onboa ding and aining. The Wo k Ins i u e (2024) es ima es ha eplacing a single employee cos s
6
app oxima ely 33% o hei annual sala y, making u no e a subs an ial inancial bu den, pa icula ly in
indus ies wi h high u no e a es such as e ail. Ope a ionally, equen employee depa u es dis up
wo k low and educe p oduc i i y, as new hi es equi e ime o adap and each peak pe o mance le els
(Rahaman & Ba i, 2024). In cus ome - acing indus ies like e ail, u no e nega i ely impac s se ice
quali y and cus ome sa is ac ion, as expe ienced employees a e mo e e ec i e in handling cus ome
in e ac ions and ope a ional demands (Ba -Gil e al., 2024).
Beyond inancial and ope a ional conce ns, u no e also has psychological and cul u al e ec s
wi hin o ganiza ions. A high u no e a e can lowe employee mo ale, inc ease wo kplace s ess and
educe o e all job engagemen . The depa u e o expe ienced employees leads o he loss o ins i u ional
knowledge, u he weakening long- e m s a egic goals and business con inui y (Gamba e al., 2024).
Consequen ly, o ganiza ions a e inc easingly ocusing on p oac i e s a egies o manage u no e
e ec i ely.
One o he mos e ec i e app oaches o mi iga ing u no e is he use o People Analy ics, which
enables HR depa men s o analyze key wo k o ce indica o s — such as job sa is ac ion, ca ee
p og ession and wo kplace sen imen — o an icipa e employee depa u es and de elop a ge ed e en ion
s a egies. P edic i e analy ics in HR allows o ganiza ions o shi om eac i e u no e managemen o
p oac i e wo k o ce planning, ensu ing business con inui y and op imizing alen e en ion (Rahaman &
Ba i, 2024). Companies ha s a egically inco po a e p edic i e modeling in o hei wo k o ce planning
p ocesses can enhance wo k o ce s abili y, educe hi ing cos s and c ea e a mo e esilien and engaged
wo k o ce (Le enson, 2018).
2.2 T adi ional Models o Employee Tu no e
The s udy o employee u no e has gained signi ican a en ion o e he yea s, leading o he
de elopmen o se e al heo e ical models ha explain why employees decide o lea e hei o ganiza ions.
Among he mos in luen ial con ibu ions in his ield a e he wo ks o Mobley (1977), Lee (1988), Lee &
Mi chell (1994), Mo ell e al. (2008) and Wöcke & Heymann (2012), each o e ing di e en pe spec i es
on u no e decision-making.
One o he ea lies and mos ounda ional models is Mobley’s amewo k, which emphasizes he
ela ionship be ween job dissa is ac ion and olun a y u no e . Acco ding o his model, employees who
expe ience dissa is ac ion begin e alua ing al e na i e job oppo uni ies, compa ing o e s and assessing
he po en ial cos s and bene i s o lea ing hei cu en ole. This pe spec i e sugges s ha he decision
7
o lea e an o ganiza ion is la gely d i en by pe cei ed job al e na i es and he weighing o isks associa ed
wi h he ansi ion (G. J. Lee & Rwigema, 2005; Wöcke & Heymann, 2012).
Building upon Mobley’s indings, Lee in oduced a e ined app oach ha shi ed he ocus om
job sa is ac ion o job commi men and in ol emen as he key de e minan s o olun a y u no e . His
esea ch highligh s ha employees’ le el o a achmen o he o ganiza ion plays a c ucial ole in
in luencing hei decision o s ay o lea e, ein o cing he idea ha commi men , a he han sa is ac ion
alone, dic a es u no e beha io (Lee, 1988).
A mo e dynamic and comp ehensi e app oach was la e p oposed by Lee & Mi chell (1994) wi h
he de elopmen o he
Un olding Model o Tu no e
. This amewo k sugges s ha u no e decisions
a e o en igge ed by ex e nal o in e nal shocks ha lead employees o eassess hei employmen
si ua ion. The model ou lines di e en decision-making pa hs, conside ing scena ios in which employees
unde go a ee alua ion p ocess ei he due o a signi ican ex e nal e en , an unexpec ed ca ee
oppo uni y o a g adual eassessmen o hei cu en ole. In some cases, employees migh be p omp ed
o lea e by a sudden ca ee dis up ion, while in o he s, u no e esul s om a long- e m decline in job
commi men wi hou an ex e nal igge .
Seeking o e ine and imp o e he p edic i e accu acy o u no e models, Mo ell e al.
in oduced he
Mapping he Decision o Qui
amewo k, which builds upon he
Un olding Model
by
emphasizing ha u no e is a ely an impulsi e decision bu a he a g adual and s uc u ed p ocess.
This model a gues ha employees ypically go h ough mul iple s ages be o e making a inal decision o
lea e, ein o cing he idea ha u no e should be seen as a se ies o e ol ing conside a ions a he han
a single e en . Compa ed o i s p edecesso s, his model is conside ed o p o ide a mo e accu a e and
s uc u ed depic ion o u no e beha io , allowing o ganiza ions o an icipa e employee depa u es wi h
g ea e p ecision (Mo ell e al., 2008).
These heo e ical models ha e signi ican ly shaped u no e esea ch by p o iding insigh s in o
he a ious ac o s ha d i e employee depa u es. While ea ly models p ima ily linked u no e o
dissa is ac ion and commi men , la e amewo ks in eg a ed ex e nal in luences, ca ee shocks and long-
e m psychological p ocesses, emphasizing he complex and mul i- ace ed na u e o u no e .
Unde s anding hese models emains essen ial o o ganiza ions seeking o de elop e ec i e e en ion
s a egies and minimize wo k o ce dis up ions.
Despi e he signi ican con ibu ions o adi ional models in explaining employee u no e , hey
also exhibi limi a ions ha educe hei p edic i e powe and p ac ical applicabili y. One o he p ima y
sho comings is hei limi ed p edic i e accu acy, as hese models o en ocus on indi idual psychological
8
ac o s such as job sa is ac ion and o ganiza ional commi men bu ail o inco po a e a comp ehensi e
se o a iables ha can enhance p edic ion accu acy (Mo ell e al., 2008). Resea ch has shown ha
u no e decisions a e in luenced by a wide a ay o ac o s, many o which a e no adequa ely cap u ed
by adi ional amewo ks. Addi ionally, adi ional models ail o conside ex e nal in luences, such as
mac oeconomic condi ions, labo ma ke ends and o ganiza ional es uc u ing, which can signi ican ly
impac an employee's decision o lea e. While he
Un olding Model
in oduced he concep o ex e nal
"shocks" igge ing u no e , ea lie models p ima ily ea ed u no e as a g adual p ocess based on
dissa is ac ion, o e looking ab up decision-making scena ios (Lee & Mi chell, 1994).
Ano he c i ical limi a ion o many adi ional u no e models is hei endency o assume linea
ela ionships be ween a iables, whe eas eal-wo ld u no e decisions o en exhibi nonlinea pa e ns
in luenced by mul iple in e ac ing ac o s (Lee & Mi chell, 1994). These models end o simpli y u no e
decision-making in o p ede ined s ages, which may no ully e lec he dynamic and e ol ing na u e o
employee exi decisions. Addi ionally, con en ional s a is ical app oaches, such as logis ic eg ession,
domina e u no e esea ch bu ace limi a ions when applied o eal-wo ld wo k o ce da a. These
me hods s uggle wi h imbalanced da ase s, whe e he numbe o employees who lea e is signi ican ly
smalle han hose who s ay, leading o biased p edic ions and educed gene alizabili y. Fu he mo e,
adi ional app oaches o en ely on p ede ined assump ions abou a iable ela ionships, which limi s
hei abili y o de ec hidden co ela ions and cap u e he in e play be ween mul iple ac o s ha in luence
u no e , pa icula ly when hese in e ac ions e ol e o e ime (Pa k e al., 2024).
Mode n app oaches le e aging machine lea ning echniques, such as decision ees and
ensemble models, ha e demons a ed supe io p edic i e pe o mance by au oma ically lea ning om
da a pa e ns a he han elying on ixed assump ions. Unlike adi ional me hods, machine lea ning
models can p ocess high-dimensional da ase s, cap u ing in ica e ela ionships be ween employee
a ibu es, wo kplace condi ions and ex e nal labo ma ke ac o s (Pa k e al., 2024). Consequen ly, as
wo k o ce planning becomes inc easingly da a-d i en, adi ional u no e models ace g owing challenges
in deli e ing ac ionable and p ecise insigh s, ein o cing he need o mo e sophis ica ed p edic i e
me hodologies.
2.3 Fac o s a ec ing Employee Tu no e
Employee u no e is in luenced by mul iple ac o s, which can be ca ego ized in o inancial,
o ganiza ional, demog aphic and economic aspec s. Unde s anding hese de e minan s is c ucial o
de eloping p edic i e models and implemen ing e en ion s a egies. While no single ac o can ully
9
explain why employees lea e, esea ch has iden i ied se e al key d i e s o u no e ha in e ac in
complex ways. Below a e ou lined he mos signi ican ac o s a ec ing employee e en ion:
• Sala y and Compensa ion: Employees who eel unde paid o pe cei e wage dispa i ies compa ed
o compe i o s a e mo e likely o lea e in sea ch o be e inancial ewa ds. Resea ch consis en ly
highligh s pay dissa is ac ion as a leading cause o olun a y u no e (Solomon e al., 2024).
• Non-Mone a y Bene i s: While sala y alone does no ensu e e en ion, bene i s such as lexible
wo k a angemen s, bonuses and heal h bene i s con ibu e o job sa is ac ion and play a ole in
employee e en ion (O ujaliye , 2024).
• Wo kload and Bu nou : Employees expe iencing excessi e wo kloads and high-p essu e
en i onmen s epo inc eased s ess, leading o disengagemen and highe u no e a es.
Indus ies like e ail, whe e job in ensi y is high, show a pa icula ly s ong co ela ion be ween
wo kload and u no e (Coombs, 2024; H. Kim & S one , 2008; Koo e al., 2020).
• Ca ee De elopmen Oppo uni ies: Employees a e mo e likely o s ay in o ganiza ions ha
p o ide s uc u ed ca ee p og ession and p o essional de elopmen p og ams. A lack o ca ee
g ow h is a s ong p edic o o olun a y u no e (Chow e al., 2007).
• Tenu e: Employees wi h sho e enu es exhibi a highe likelihood o lea ing, whe eas hose wi h
longe enu es end o s ay unless hey pe cei e a lack o ca ee ad ancemen (Ju & Li, 2019).
• Job Sa is ac ion: Employees dissa is ied wi h hei daily esponsibili ies, wo kplace en i onmen
o in e pe sonal ela ionships a e mo e p one o lea ing. Poo leade ship and ine ec i e
communica ion con ibu e o highe u no e , whe eas a posi i e wo kplace cul u e educes hese
isks (Judge & Kammeye -Muelle , 2012; E. G. Kim & Kim, 2021; Tian e al., 2020).
• Leade ship and O ganiza ional Cul u e: Leade ship s yle and wo kplace cul u e signi ican ly
in luence employee e en ion. Suppo i e leade ship, anspa en communica ion and a cul u e
o ecogni ion con ibu e o a posi i e wo k en i onmen , os e ing employee commi men . In
con as , poo leade ship, lack o manage ial suppo and a oxic wo k cul u e inc ease u no e
isks, as employees may eel unde alued o unmo i a ed (Bass, 1990; Tian e al., 2020).
• Demog aphics (Age, Gende and Educa ion Le el):
o Age: Younge employees end o seek new oppo uni ies, while olde employees p e e
s abili y (Ng & Feldman, 2009).
o Gende : Resea ch sugges s male employees a e mo e likely o lea e han emale
employees, who may p io i ize job s abili y (G issom e al., 2012).
16
accu acy, a 90.79% F1 sco e, 97.18% p ecision and 85.19% ecall. GBT also pe o med well, wi h an F1
sco e o 87.42% and ecall o 81.48%. These esul s highligh he supe io pe o mance o XGBoos in
p edic ing employee u no e , making i he mos sui able model o his ask, ollowed closely by G adien
Boos ing T ee.
The s udy by Punnoose & Aji (2016) aims o imp o e he p edic ion o employee u no e using
machine lea ning echniques, wi h a ocus on XGBoos . The backg ound o he s udy emphasizes ha
employee u no e is a signi ican issue o o ganiza ions, impac ing p oduc i i y and business con inui y.
Accu a e u no e p edic ion helps companies implemen mo e e ec i e e en ion s a egies. Howe e ,
he main challenge lies in he quali y o he da a used, which o en con ains noise due o insu icien
in es men in Human Resou ce In o ma ion Sys ems (HRIS). Despi e his, his model demons a ed
excellen esul s, ou pe o ming all o he models, including Logis ic Reg ession, Random Fo es and Naï e
Bayes. The s udy ound ha XGBoos achie ed an AUC sco e o 0.86 and showed high accu acy in
p edic ing u no e . Fu he mo e, i excelled in e ms o low memo y usage and as aining imes, which
makes i highly e icien o u no e p edic ion asks. This s udy, based on da a om a global e aile ,
con i ms he obus ness o XGBoos , solidi ying i as a s ong choice o u no e o ecas ing, pa icula ly
in noisy da a scena ios.
Simila indings we e epo ed by Zhao e al. (2019), who expanded he scope by es ing
addi ional models, including Neu al Ne wo ks (NNs) and Suppo Vec o Machines (SVMs), ein o cing
he e iciency o XGBoos in handling la ge da ase s. In his a icle, a comp ehensi e e alua ion o
supe ised machine lea ning algo i hms o p edic employee u no e ac oss small, medium and la ge-
scale o ganiza ional da ase s was conduc ed. The s udy examined mul iple models, including Decision
T ees, Random Fo es s, G adien Boos ing T ees, Ex eme G adien Boos ing (XGBoos ), Logis ic
Reg ession, Suppo Vec o Machines, Neu al Ne wo ks, Linea Disc iminan Analysis, Naï e Bayes and
K-Nea es Neighbo s. The indings highligh ed he supe io p edic i e pe o mance o XGBoos and
G adien Boos ing T ees, pa icula ly o medium and la ge da ase s, due o hei abili y o handle complex
in e ac ions, educe o e i ing h ough egula iza ion and ank ea u e impo ance au oma ically.
Addi ionally, he s udy ound ha XGBoos ou pe o med adi ional G adien Boos ing Machines (GBM)
in e ms o compu a ional e iciency, as i equi ed sho e aining imes while main aining high p edic i e
accu acy. This ad an age was a ibu ed o XGBoos ’s op imized pa alleliza ion, memo y-e icien ee
s uc u e and in eg a ed egula iza ion echniques. The s udy also emphasized ha small da ase s o en
in oduce high a iance and andomness, making model selec ion less eliable in hose cases.
Fu he mo e, Zhao e al. unde sco ed he impo ance o app op ia e da a p ep ocessing and model
17
in e p e abili y, ad oca ing o ea u e anking and classi ica ion ule isualiza ion o enhance he p ac ical
applica ion o u no e p edic ion models.
The s udy by Pa k e al. (2024) e alua ed he pe o mance o mul iple machine lea ning models
in p edic ing employee u no e in en ion, highligh ing he ad an ages o XGBoos o e adi ional
s a is ical app oaches. Among he es ed models, XGBoos demons a ed he highes accu acy (78.5%),
ou pe o ming Logis ic Reg ession (78.3%) and K-Nea es Neighbo s (KNN) (76.1%). A key s eng h o
XGBoos no ed in he s udy was i s abili y o ank ea u e impo ance e ec i ely, allowing o ganiza ions o
iden i y he mos in luen ial ac o s d i ing u no e . No ably, job secu i y eme ged as he mos c i ical
p edic o o u no e in en ion, ein o cing he ole o wo kplace s abili y in employee e en ion.
Fu he mo e, he s udy in eg a ed machine lea ning echniques wi h adi ional econome ic analysis,
demons a ing how ad anced p edic i e models can enhance wo k o ce planning by p o iding da a-d i en
insigh s in o employee beha io .
Table 1 p esen s a compa a i e summa y o he me hodologies used in he e iewed s udies. I
highligh s he machine lea ning models e alua ed, he pe o mance me ics conside ed, he alida ion
s a egies applied and he speci ic con ex s in which hese models we e es ed.
18
Table 1 - O e iew o Tu no e P edic ion S udies: Models, Me ics and Applica ions
As summa ized in
Table 1
, all e iewed s udies included XGBoos in hei compa isons,
consis en ly demons a ing i s supe io p edic i e pe o mance o e models such as Random Fo es ,
Logis ic Reg ession and Naï e Bayes. While he choice o e alua ion me ics a ied, accu acy and AUC
A icle
Models Compa ed
E alu ion Me ics
E alua ion S a egy
Applica ion Con ex
Ko u i &
Domme i
(2022)
Logis ic
Reg ession, Naï e
Bayes, Random
Fo es , XGBoos
Accu acy,
P ecision, Recall,
F1 Sco e,
AUC-ROC
75/25
T ain-Tes Spli
Employee Tu no e
P edic ion using Kaggle
da ase
Ju i ayapun
(2021)
Logis ic
Reg ession,
Random Fo es ,
G adien Boos ing
T ees, XGBoos
Accu acy,
P ecision, Recall,
F1 Sco e,
ROC AUC,
P edic ion Cos
10- old C oss-
Valida ion,
G id Sea ch o
Hype pa ame e
Tuning
Employee Tu no e
P edic ion in a Packaging
Manu ac u ing Fi m
(Thailand)
Punnoose &
Aji (2016)
XGBoos , Logis ic
Reg ession, Naï e
Bayes, Random
Fo es , SVM, LDA,
KNN
AUC-ROC,
Run ime,
Memo y
U iliza ion
10- old C oss-
Valida ion,
80/20
T ain-Tes Spli
Employee Tu no e
P edic ion in a Global
Re aile (US wo k o ce)
Zhao e al.
(2019)
Decision T ee,
Random Fo es ,
G adien Boos ing
T ees, XGBoos ,
Logis ic
Reg ession, SVM,
Neu al Ne wo ks,
LDA, Naï e Bayes,
KNN
Accu acy,
P ecision, Recall,
F1 Sco e,
AUC-ROC
10- old C oss-
Valida ion,
G id Sea ch o
Hype pa ame e
Tuning
Employee Tu no e
P edic ion using da ase s
om a US Bank and IBM
Wa son Analy ics
Pa k e al.
(2024)
Logis ic
Reg ession,
K-Nea es
Neighbo s,
XGBoos
Accu acy,
P ecision, Recall,
F1 Sco e
70/30
T ain-Tes Spli ,
4- old C oss-
Valida ion
P edic ion o u no e
in en ion among new
college g adua es in
Sou h Ko ea
19
we e he mos commonly epo ed. Addi ionally, alida ion s a egies di e ed ac oss s udies, wi h some
employing k- old c oss- alida ion and o he s elying on simple holdou alida ion, emphasizing he need
o s anda dized benchma king app oaches in employee u no e p edic ion esea ch.
O e all, hese s udies ein o ce XGBoos 's e ec i eness in employee u no e p edic ion, as i
consis en ly ou pe o ms adi ional models ac oss di e se da ase s and alida ion se ings.
2.5.2 XGBoos : A Benchma k Model o Tu no e P edic ion
XGBoos is a machine lea ning model based on he G adien Boos ing amewo k, widely ecognized
o i s e iciency and abili y o handle la ge da ase s e ec i ely. In he con ex o u no e p edic ion,
XGBoos is pa icula ly aluable due o i s s ong pe o mance in handling complex, high-dimensional and
noisy da a while main aining bo h accu acy and compu a ional e iciency (Punnoose & Aji , 2016).
XGBoos cons uc s decision ees sequen ially, wi h each new ee lea ning om and co ec ing
he e o s made by he p e ious ees. This p ocess, known as boos ing, enhances p edic i e accu acy
by ocusing on he esiduals — he di e ences be ween he ac ual and p edic ed alues. Addi ionally,
XGBoos inco po a es bo h L1 and L2 egula iza ion echniques, which a e c ucial o p e en o e i ing,
pa icula ly in u no e p edic ion da ase s whe e many ea u es and po en ial collinea i y could o he wise
comp omise model gene aliza ion (T. Chen & Gues in, 2016).
A key ad an age o XGBoos is i s scalabili y, as i op imizes bo h ime and memo y usage, enabling
i o e icien ly p ocess la ge olumes o da a (T. Chen & Gues in, 2016). This scalabili y is especially
impo an in u no e p edic ion scena ios, whe e da a o en spans se e al yea s and includes nume ous
ea u es cap u ing bo h indi idual employee cha ac e is ics and o ganiza ional ac o s (Punnoose & Aji ,
2016). In addi ion, XGBoos handles missing da a na i ely, au oma ically lea ning he op imal spli
di ec ion o eco ds wi h missing alues, ensu ing hey a e e ec i ely inco po a ed in o he aining
p ocess wi hou equi ing p io impu a ion (T. Chen & Gues in, 2016). Ano he impo an ea u e o
XGBoos is i s abili y o handle imbalanced da a, which is common in u no e p edic ion, whe e u no e
e en s ypically ep esen a smalle p opo ion o he o al da ase . The pa ame e
scale_pos_weigh
can
be adjus ed o gi e highe impo ance o he mino i y class ( u no e ), ensu ing he model emains
sensi i e o hese c i ical cases (T. Chen & Gues in, 2016; Punnoose & Aji , 2016). Fu he mo e, XGBoos
o e s ex ensi e lexibili y h ough i s hype pa ame e uning op ions, including
lea ning a e
,
max_dep h
,
subsample
and
n_es ima o s
. These pa ame e s con ol he lea ning p ocess, ee complexi y and da a
sampling s a egy, allowing he model o be ca e ully op imized o maximum p edic i e accu acy (T. Chen
& Gues in, 2016).
20
O e all, his chap e has explo ed he concep ual ounda ions and ecen ad ances in employee
u no e p edic ion, emphasizing he impo ance o inco po a ing bo h in insic and ex insic ac o s in o
wo k o ce planning. I also highligh ed he limi a ions o adi ional s a is ical app oaches and he
ad an ages o machine lea ning echniques — pa icula ly XGBoos — in cap u ing complex, non-linea
pa e ns in employee da a. These insigh s p o ide a solid basis o he me hodological app oach adop ed
in his disse a ion. The nex chap e will con ex ualize hese indings by examining he company’s cu en
decision-making p ocess and de ailing he da ase used o build he p oposed p edic i e model.
2.6 Summa y o Key Findings
The li e a u e e iewed in his chap e highligh s se e al c i ical insigh s ele an o he
de elopmen o a u no e p edic ion model in he e ail sec o . Fi s , employee u no e is a mul i ace ed
phenomenon in luenced by bo h in insic ac o s — such as compensa ion, pe o mance, senio i y and
job sa is ac ion — and ex insic ac o s, including labo ma ke ends and o ganiza ional cul u e.
T adi ional models o u no e , while aluable o unde s anding indi idual decision-making p ocesses,
a e limi ed in hei p edic i e accu acy and scalabili y. In con as , he g owing use o People Analy ics
and machine lea ning echniques has demons a ed signi ican po en ial o imp o ing u no e p edic ion
accu acy and enabling p oac i e wo k o ce planning. XGBoos , in pa icula , s ands ou due o i s abili y
o handle imbalanced and high-dimensional da a, pe o m egula iza ion and p o ide in e p e able ea u e
impo ance ankings. Fu he mo e, p edic i e analy ics no only enhances s a egic HR decision-making
bu also con ibu es o cos e iciency, e en ion planning and long- e m wo k o ce s abili y.
Gi en he ola ili y o he e ail wo k o ce and he ope a ional challenges posed by high u no e
a es, he li e a u e also emphasizes he need o obus , da a-d i en decision-suppo ools capable o
add essing his complexi y. These indings unde sco e he alue o in eg a ing ad anced analy ical models
in o HR p ac ices and p o ide he concep ual ounda ion o he me hodology adop ed in his disse a ion.
21
3. THE PROBLEM
Following he li e a u e e iew, which highligh ed bo h he heo e ical ounda ions and mode n
ad ances in u no e p edic ion, his chap e in oduces he eal-wo ld con ex in which he p oposed
solu ion will be applied. The aim is o b idge he gap be ween academic models and p ac ical
implemen a ion by de ailing he p oblem ha mo i a ed he de elopmen o he p edic i e amewo k.
In pa icula , his chap e begins by desc ibing he cu en decision-making p ocess used by he
company o es ima e olun a y u no e , highligh ing i s empi ical na u e and inhe en limi a ions. I hen
p o ides an o e iew o he a ailable da a, including i s s uc u e, key a iables and sou ces. Finally, an
explo a o y da a analysis is conduc ed o iden i y ends, alida e da a quali y and suppo he design o
he p edic i e model in he nex chap e .
3.1 P oblem Desc ip ion
3.1.1 As-Is
In s a egic wo k o ce planning, ha ing a eliable app oach o an icipa ing wo k o ce needs is essen ial
o ensu ing ope a ional e iciency. While well-s uc u ed wo k o ce plans help mi iga e s a ing isks,
na u al employee u no e emains an inhe en challenge, making i c ucial o in eg a e wo k o ce
p ojec ions in o decision-making p ocesses. By p oac i ely es ima ing u u e employee depa u es,
companies can ensu e a balanced wo k o ce s uc u e, educing ine iciencies and op imizing ec ui men
e o s.
Cu en ly, wo k o ce adjus men s a e p ima ily in o med by his o ical wo k o ce ends, using pas
a ia ions o app oxima e u no e expec a ions. One o he key e e ences in his p ocess is he FTE-
based me hod, which moni o s u no e pa e ns o e ecen yea s o p o ide insigh s in o wo k o ce
dynamics. While his app oach o e s a p ac ical benchma k o wo k o ce planning, i does no
inco po a e indi idual employee cha ac e is ics o ex e nal labo ma ke condi ions, making i less
adap able o e ol ing wo k o ce ends.
In addi ion o his o ical ends, he company also u ilizes he Value-Based Managemen (VBM) sys em,
which p o ides de ailed wo k o ce p ojec ions pe ole clus e , s o e and yea . These p ojec ions help
align s a ing needs wi h expec ed wo k o ce changes, ensu ing ha adjus men s can be made p oac i ely
a he han eac i ely.
As wo k o ce dynamics become inc easingly complex, enhancing wo k o ce planning wi h p edic i e
models p esen s an oppo uni y o e ine u no e es ima ions, in eg a ing bo h in insic (e.g., employee
22
a ibu es) and ex insic (e.g., ma ke condi ions) ac o s. This app oach aims o complemen exis ing
wo k o ce planning s a egies, allowing o a mo e adap i e and da a-d i en decision-making p ocess.
3.1.2 Poin o Imp o emen
Despi e hese s uc u ed p ocesses, he e a e wo key a eas whe e imp o emen s could
signi ican ly enhance he company’s abili y o accu a ely p edic and manage wo k o ce changes, he eby
os e ing a mo e s a egic and e icien wo k o ce planning app oach.
Fi s , he cu en u no e p edic ion me hod does no inco po a e employee- ela ed a iables,
such as job sa is ac ion, enu e, pe o mance e alua ions, con ac ype, wo k schedule and ex e nal labo
ma ke condi ions. These in insic and ex insic ac o s a e key d i e s o olun a y u no e , in luencing
employees' decisions o s ay o lea e. Howe e , hey a e no cap u ed in he FTE-based es ima ion model,
which solely elies on his o ical wo k o ce a ia ions. This limi a ion educes he company’s abili y o
an icipa e wo k o ce gaps wi h p ecision, making i di icul o implemen p oac i e e en ion s a egies,
op imize s a ing le els and mi iga e ope a ional dis up ions. As a esul , he company emains in a
eac i e s ance, whe e adjus men s o wo k o ce planning a e made a e u no e occu s a he han in
an icipa ion o i .
Second, he company’s ini ial wo k o ce needs a e de e mined by Senio People Manage s based
p ima ily on expe ience and manage ial in ui ion, a he han on s uc u ed, da a-d i en me hodologies.
While his expe ise is aluable, i in oduces an elemen o subjec i i y, leading o po en ial
inconsis encies in wo k o ce alloca ion ac oss di e en s o e o ma s, egions and job oles. This eliance
on human judgmen may esul in o e s a ing in some a eas, leading o unnecessa y labo cos s o
unde s a ing in c i ical oles, impac ing ope a ional e iciency and se ice le els. Fu he mo e, his
manual decision-making p ocess demands a signi ican ime in es men om manage s, which could
o he wise be alloca ed o mo e s a egic wo k o ce ini ia i es.
By in eg a ing employee-speci ic a iables in o u no e p edic ion models and ansi ioning
owa ds a da a-d i en app oach in wo k o ce planning, he company can gain a mo e accu a e, g anula
unde s anding o he ac o s d i ing employee depa u es. This shi would enable p oac i e wo k o ce
managemen , whe e HR eams can an icipa e u no e isks, adjus hi ing s a egies acco dingly and
implemen a ge ed e en ion e o s o mi iga e unnecessa y a i ion. Addi ionally, le e aging p edic i e
analy ics in wo k o ce planning would es ablish a mo e consis en and objec i e me hodology, educing
eliance on manage ial in ui ion and ensu ing ha s a ing decisions a e aligned wi h eal ope a ional
demands.
23
Implemen ing hese imp o emen s would lead o mul iple bene i s, including:
• Op imized esou ce alloca ion, ensu ing ha each s o e o ma has he necessa y wo k o ce
capaci y.
• Cos educ ion, by minimizing unexpec ed hi ing needs and excessi e labo cos s.
• G ea e wo k o ce s abili y, educing ope a ional dis up ions caused by unexpec ed employee
depa u es.
• Enhanced s a egic decision-making, allowing HR and managemen eams o ocus on long- e m
wo k o ce planning a he han eac i e adjus men s.
By e ining i s u no e p edic ion app oach and adop ing a mo e s uc u ed, analy ics-based
wo k o ce planning model, he company can enhance ope a ional e iciency, educe wo k o ce
managemen cos s and ensu e sus ainable business g ow h in an inc easingly compe i i e e ail
landscape.
3.2 Da a O e iew and Explo a o y Analysis
3.2.1 Da a O e iew
The da ase used o his analysis comp ises a o al o 189,160 employee eco ds pe yea , spanning
om 2015 o 2023. I is o ganized by "employee x yea ", meaning ha o each yea , he e is a eco d
o e e y employee, p o iding a obus and comp ehensi e longi udinal iew o he wo k o ce o e ime.
The da ase consis s o a ious a iables ha cap u e di e en aspec s o employee cha ac e is ics,
u no e and wo k- ela ed ac o s. These a iables include:
• Employee ID: A unique iden i ie o each employee.
• Yea : The yea o which he employee eco d co esponds.
• Tu no e : Indica ing whe he he employee has le he o ganiza ion.
• Sala y: The ixed sala y paid o he employee.
• S o e Type: The ype o s o e whe e he employee wo ks.
• Role Clus e : The job ole classi ica ion o he employee.
• Academic Quali ica ions: The educa ional backg ound o he employee.
• Con ac Type: Type o employmen con ac (e.g., pe manen , ixed e m).
• Ma i al S a us: Employee's ma i al s a us.
• Age: The employee's age.
• Senio i y: The leng h o ime he employee has wo ked wi h he company.
24
• Wo kload Type: The ype o wo kload (e.g., ull- ime, pa - ime).
The da ase is gene ally well-popula ed, wi h he highes p opo ion o missing alues obse ed in he
wo kload a iable, which accoun s o 10% o he o al da a. Fo all o he a iables, missing da a is
minimal, wi h an a e age o 1.5% missing alues ac oss he da ase . This low le el o missing da a ensu es
ha he analysis can be conduc ed wi h a high deg ee o comple eness.
Ou lie s we e iden i ied based on he in e qua ile ange (IQR) me hod, which allows o de ec ing
ex eme alues in nume ical da a. The a iable wi h he highes pe cen age o ou lie s is sala y, a 11%.
This is likely due o a small p opo ion o employees in highe -le el posi ions who ecei e signi ican ly
highe sala ies compa ed o he majo i y o employees in lowe -le el oles. These ou lie s a e conside ed
na u al and e lec eal di e ences in pay scales wi hin he company. O e all, 6.4% o he da ase con ains
ou lie s ac oss all a iables.
3.2.2 Da a Analysis
In his sec ion, we p esen an analysis o key a iables and hei ela ionships wi hin he da ase .
The goal is o unde s and he pa e ns, ends and dis ibu ions ha will help us build a mo e e ec i e
p edic i e model o employee u no e . This ounda ional unde s anding will guide he modeling
app oach by highligh ing impo an ac o s ha need o be conside ed.
The Figu e 1 illus a es he annual e olu ion o he wo k o ce and u no e a e. The main end
obse ed is he g adual g ow h o he numbe o employees o e ime, excep du ing he
COVID
yea s
(2020 and 2021). Du ing hese yea s, he u no e a e d ops signi ican ly, likely due o he economic
unce ain y ha made employees mo e inclined o e ain hei cu en jobs due o he ins abili y o he
job ma ke . Also, he inc easing end in he numbe o employees e lec s he company's expansion,
which is c ucial o o ecas ing u no e and s a ing needs in he coming yea s.
Figu e 1 - Annual Wo k o ce E olu ion and Tu no e Ra e
25
Figu e 2 highligh s he pe cen age o employees who le he company compa ed o hose who
s ayed. I shows ha , on a e age, a ound 10.3% o employees le , while 89.7% emained. This imbalance
is c ucial o p edic i e modeling, as unbalanced da a like his (whe e he numbe o depa u es is
signi ican ly lowe han he numbe o employees emaining) can lead o biased p edic ions and educed
model pe o mance, pa icula ly by a ec ing he model’s abili y o lea n om he mino i y class
(employees who le he company). To imp o e he accu acy o u no e p edic ions, i is essen ial o
add ess his imbalance using echniques such as o e sampling, unde sampling o adjus ing he cos
unc ion o gi e mo e weigh o he mino i y class.
Figu e 2 - Pe cen age o Employees wi h and wi hou Tu no e
Figu e 3 shows he wo k o ce dis ibu ion and co esponding u no e a es ac oss he en
dis ic s wi h he highes employee ep esen a ion, which oge he accoun o app oxima ely 87.5% o
he o al wo k o ce. Lisbon and Po o s and ou , ep esen ing 28.02% and 21.25% o employees,
espec i ely.
Tu no e a es ac oss hese dis ic s gene ally ange be ween 9% and 12%, wi h a ew excep ions.
No ably, Po o, despi e being he second-la ges employing dis ic , eco ds he second-lowes u no e
a e (8.35%), anking 9 h among he en in e ms o employee a i ion. This con as may poin o g ea e
job s abili y o s onge employee-company alignmen in he egion.
The Madei a egion p esen s a pa icula ly low u no e a e o 4.56%, well below all o he
dis ic s. This di e ence may be pa ially explained by i s geog aphic cha ac e is ics — being an island —
which could educe ex e nal job oppo uni ies and con ibu e o g ea e wo k o ce e en ion.
32
wi h ull- ime employees ha ing a ela i ely low u no e a e o 5.97%, while pa - ime employees exhibi
a much highe u no e a e o 19.49%.
This dispa i y sugges s ha pa - ime employees a e mo e likely o lea e he company compa ed
o hei ull- ime coun e pa s. Se e al ac o s could con ibu e o his end, including he ypically lowe
job secu i y, ewe ca ee ad ancemen oppo uni ies and a g ea e likelihood o pa - ime wo ke s using
hese posi ions as ansi ional oles be o e mo ing on o ull- ime jobs elsewhe e. Addi ionally, pa - ime
oles o en a ac s uden s o indi iduals seeking supplemen a y income, making hem mo e p one o
olun a y u no e as hei ci cums ances change.
On he o he hand, ull- ime employees end o ha e g ea e job s abili y, possibly bene i ing om
be e compensa ion, bene i s and ca ee p og ession oppo uni ies, which may con ibu e o a lowe
u no e a e. This insigh highligh s he need o dis inc e en ion s a egies o pa - ime employees,
especially i educing u no e wi hin his g oup is a p io i y o he o ganiza ion.
Figu e 40 p o ides addi ional insigh s in o he in e ac ion be ween wo kload ype and con ac
s abili y, illus a ing how pa - ime and ull- ime oles a e dis ibu ed ac oss di e en con ac ypes. Gi en
he u no e pa e ns obse ed in Figu e 8 and Figu e 11, unde s anding his ela ionship may help cla i y
whe he con ac ype and wo kload ac as o e lapping ac o s in luencing employee u no e .
Figu e 11 - Wo k o ce Dis ibu ion and Tu no e Ra e by Wo kload Type
Figu e 12 shows he dis ibu ion o employees and co esponding u no e a es by pe o mance le el
( a ed on a scale om 1 o 5). The as majo i y o employees — o e 80% — a e e alua ed wi h a
pe o mance sco e o 3, making i he dominan ca ego y wi hin he o ganiza ion. This cen al endency
may e lec a conse a i e o s anda dized e alua ion cul u e.
Tu no e a es ollow a cu ed end, peaking a he mid-pe o mance le el (3) wi h 20.81% and
dec easing as pe o mance mo es owa d ei he ex eme. Employees a ed as 4 o 5 exhibi lowe
33
u no e (12.59% and 7.63%, espec i ely), which aligns wi h expec a ions, as high-pe o ming indi iduals
a e ypically mo e engaged o may bene i om e en ion e o s. On he o he end, al hough employees
a ed 1 o 2 ep esen a minimal ac ion o he popula ion, hei u no e a es a e non-negligible (9.39%
and 16.25%, espec i ely), which could indica e some le el o olun a y o in olun a y a i ion among
lowe pe o me s.
The ac ha mos e alua ions clus e a le el 3, wi h de ia ions owa d 1 o 5 occu ing only in a
mino i y o cases, sugges s ha pe o mance a ings a e only signi ican ly adjus ed when an employee is
pe cei ed as ei he clea ly unde pe o ming o excelling. As such, sco es a he ex emes may se e as
s ong indica o s o key wo k o ce decisions, such as p omo ions o exi p io i iza ion.
Figu e 12 - Wo k o ce Dis ibu ion and Tu no e Ra e by Employee Pe o mance
Figu e 13 p esen s he dis ibu ion o employees and co esponding u no e a es by po en ial le el
( a ed on a scale om 1 o 3). The majo i y o employees — nea ly 80% — a e a ed wi h a po en ial sco e
o 1, ollowed by 16% a le el 2 and jus unde 5% a le el 3. This skewed dis ibu ion sugges s ha highe
po en ial a ings a e ese ed o a ela i ely small po ion o he wo k o ce, possibly e lec ing s ic e
e alua ion c i e ia o a conse a i e app oach in alen iden i ica ion.
Tu no e a es show a clea dec easing end as po en ial inc eases. Employees wi h a po en ial
a ing o 1 exhibi a u no e a e o 11.54%, which d ops o 5.47% o le el 2 and 4.12% o le el 3. This
downwa d pa e n sugges s ha highe - a ed employees a e mo e likely o s ay wi hin he o ganiza ion,
po en ially due o s onge engagemen , g ow h oppo uni ies o e en ion ini ia i es.
The a i y o le els 2 and 3 highligh s hei possible ole in iden i ying key alen segmen s o
succession planning o ca ee de elopmen . As such, he po en ial a iable may be pa icula ly use ul in
di e en ia ing
pe sonas
wi hin he op imiza ion model, helping o in o m s a egic decisions such as
p omo ions, ole ealloca ions and a ge ed e en ion e o s.
34
Figu e 13 - Wo k o ce Dis ibu ion and Tu no e Ra e by Employee Po en ial
To u he suppo hese indings, addi ional analyses we e conduc ed on he in e ac ions be ween
enu e, con ac ype, wo kload and sala y dis ibu ion. These complemen a y insigh s can be ound in
Appendix A – Suppo ing Da a and Ex ended Analysis.
3.3 Final Rema ks
This chap e ou lines he co e p oblem add essed in his disse a ion — he high le els o
employee u no e wi hin he company and he limi ed e ec i eness o cu en heu is ic-based
app oaches in accu a ely o ecas ing u u e wo k o ce needs. The explo a o y da a analysis e ealed a
ich and g anula da ase , wi h a iables ha o e signi ican po en ial o enhance u no e p edic ion,
despi e ce ain challenges such as class imbalance and s uc u al a iabili y ac oss s o es, oles and
egions.
Fu he mo e, he dual na u e o he p edic i e ask — indi idual-le el o ecas ing o he i s yea
and agg ega e-le el o ecas ing o subsequen yea s — ein o ces he need o lexible and obus models
ailo ed o di e en planning ho izons. These indings unde sco e he limi a ions o he cu en o ecas ing
me hodology and p o ide a s ong a ionale o adop ing a da a-d i en machine lea ning app oach, which
will be de ailed in he ollowing chap e .
35
4. METHODOLOGY FOR TURNOVER PREDICTION
The objec i e o his chap e is o desc ibe he me hodology used o de elop a u no e p edic ion
model ha o ecas s he olun a y depa u e o employees o e a i e-yea ho izon. Al hough he p edic i e
model was designed wi h po en ial in eg a ion in o a b oade wo k o ce op imiza ion amewo k, his
chap e — and he disse a ion as a whole — ocuses exclusi ely on he de elopmen , implemen a ion and
e alua ion o he p edic i e componen . In olun a y exi s, such as layo s o con ac expi a ions, a e
excluded om he scope, as hey a e d i en by di e en o ganiza ional dynamics.
Two p edic i e models we e de eloped, each add essing a di e en ime ho izon. The i s model
o ecas s u no e o he ollowing yea based on he cu en cha ac e is ics o each employee. In
con as , he second model es ima es u no e o e he subsequen ou yea s (i.e., om he second o
he i h yea ), p o iding agg ega ed o ecas s by
pe sona
, ole clus e , s o e ype and egion a he han
indi idual-le el p edic ions. Since he cha ac e is ics o u u e employees a e no ye known, his second
model equi ed a s uc u al adjus men . To manage his unce ain y, a
pe sona
-based app oach was
adop ed, allowing he model o gene alize he expec ed beha io o u u e hi es based on ep esen a i e
employee p o iles.
Pe sonas
a e de ined by a combina ion o a iables: age, enu e, academic quali ica ions, con ac
ela ionship ( ixed- e m o pe manen ), wo kload ype ( ull- ime o pa - ime), pe o mance (scale o 1 o
5) and po en ial (scale o 1 o 3). Each
pe sona
is also associa ed wi h a speci ic dis ic x s o e ype x
ole clus e combina ion, ensu ing ha employee p o iles a e ailo ed o he o ganiza ional and ope a ional
con ex o each business uni . These g oupings we e de ined by he company o e lec in e nal policies,
pe o mance h esholds and de elopmen po en ial.
Fo example,
Pe sona 1
migh ep esen employees aged unde 25, wi h seconda y educa ion, a
ixed- e m con ac , wo king pa - ime and a ed wi h a pe o mance o 4 and a po en ial o 2. Acco ding
o company s a egy, his
pe sona
may be conside ed eligible o e ical mobili y (p omo ion), la e al
mo emen (changing oles) o e en lagged o exi i hey do no align wi h long- e m o ganiza ional needs.
The exac alue anges o nume ical a iables (such as age, pe o mance and po en ial), as well as he
speci ic ca ego ies o g oupings o ca ego ical a iables (such as con ac ype o academic
quali ica ions), we e de ined by he company acco ding o in e nal c i e ia and s a egic objec i es.
Al hough
pe sonas
a e no di ec ly used in he u no e p edic ion model, hey a e c i ical in he
subsequen op imiza ion phase. This op imiza ion model uses he
pe sona
s uc u e o de e mine which
p o iles a e p io i ized o p omo ion, mobili y o exi , aligning wo k o ce ansi ions wi h s a egic planning
objec i es.
36
Addi ionally, he p edic ion will be made yea by yea , conside ing only he inal s a us o each
employee a he end o each yea . In e nal dynamics ha occu wi hin each yea , such as empo a y
employee mo emen s o ole adjus men s, will no be conside ed in he p edic ions. Excluding hese
in e nal dynamics allows he model o main ain a simpli ied s uc u e, ocusing on he inal s a us o
employees, which has a mo e di ec impac on e en ion s a egies and eplacemen cos s. In his way,
he model becomes mo e e icien and ailo ed o he company’s long- e m needs, while also p o iding
mo e s able and accu a e p edic ions o suppo s a egic decision-making. Mo eo e , o align he
p edic i e ou pu s wi h he equi emen s o he wo k o ce op imiza ion amewo k, he indi idual-le el
p edic ions om he one-yea model a e subsequen ly agg ega ed by
pe sona
, ole clus e , s o e ype and
egion.
Gi en he la ge olume o da a gene a ed by he clien company, a decision ee-based app oach was
chosen. This ype o model is pa icula ly sui able o handling he e ogeneous da a, including bo h
ca ego ical and nume ical a iables and o p o iding anspa en , in e p e able ou pu s, which is c ucial
in human esou ce managemen con ex s.
Among decision ee-based models, XGBoos was selec ed due o i s e iciency and boos ing
op imiza ion capabili ies. XGBoos is pa icula ly aluable when wo king wi h la ge-scale, high-dimensional
and complex da a, whe e i balances p edic i e powe wi h compu a ional e iciency.
As discussed in he Da a O e iew and Explo a o y Analysis sec ion, he da ase con ains nea ly
200,000 eco ds, co e ing se e al yea s and including nume ous ea u es ha cap u e bo h employee
cha ac e is ics and o ganiza ional ac o s. The analysis also iden i ied he p esence o some na u al
ou lie s, pa icula ly in sala y a iables, e lec ing genuine di e ences be ween employee p o iles (e.g.,
manage ial s ope a ional oles). These ou lie s we e e ained o p ese e he ep esen a i eness o he
da ase .
Al hough boos ing algo i hms, including XGBoos , can be sensi i e o ou lie s in ce ain con ex s, ee-
based models a e gene ally less a ec ed by ou lie s in inpu ea u es han models elying on dis ance-
based calcula ions (e.g., linea eg ession o k-nea es neighbo s). This is because decision ees spli
da a based on h esholds, meaning ex eme alues a e ea ed he same as any o he alue alling in o
he same side o he spli . This cha ac e is ic, combined wi h XGBoos ’s egula iza ion mechanisms, helps
mi iga e he po en ial in luence o ou lie s on he inal model.
37
4.1 Da a P epa a ion
4.1.1 Da a Sou ces
The da a used in his s udy we e ex ac ed om he company’s in e nal Human Resou ces pla o ms.
These pla o ms eco d de ailed in o ma ion abou employees, including da a on hei ole, pe o mance,
compensa ion and his o y. These pla o ms p o ide con inuously upda ed da a, ensu ing an accu a e and
cu en pic u e o he wo k o ce.
Addi ionally,
pe sonas
we e de ined by he company as s a egic g oupings o employees based on
ole clus e s, s o es o egions, conside ing speci ic c i e ia such as pe o mance, po en ial and eligibili y
o mobili y. These
pe sonas
allow he company o be e plan wo k o ce ansi ions and op imize s a ing
decisions. Though he
pe sonas
do no di ec ly eed in o he u no e p edic ion model, hey se e as a
basis o u u e wo k o ce planning and he op imiza ion model ha will be applied a e he u no e
p edic ion.
Ex e nal a iables, such as unemploymen a e, in la ion a e, a e age wage a ia ion and minimum
wage a ia ion, we e ob ained om eliable public sou ces, such as he Na ional S a is ics Ins i u e (INE).
These a iables we e included o accoun o mac oeconomic ac o s ha may in luence employee
u no e decisions.
4.1.2 Da a In eg a ion and Consolida ion
The da a om a ious sou ces we e in eg a ed in o a mas e able, whe e each ow ep esen s an
employee in a gi en yea , iden i ied by he key "employee x yea ". This s uc u e allows acking each
employee’s his o y h oughou he analysis pe iod, acili a ing he applica ion o p edic i e models.
Table 2 - Da ase S uc u e o Tu no e P edic ion
employee code
yea
p edic o
1
p edic o 2
…
p edic o
n-1
p edic o
n
u no e
100001
2015
p ed_1_ 1
p ed_2_ 1
…
p ed_n-1_ 1
p ed_n_ 1
0
100001
2016
p ed_1_ 2
p ed_2_ 2
…
p ed_n-1_ 2
p ed_n_ 2
1
100002
2015
p ed_1_ 3
p ed_2_ 3
…
p ed_n-1_ 3
p ed_n_ 3
1
• Employee Code: Unique iden i ie o each employee.
• Yea : Yea o which he da a we e eco ded.
• P edic o s: These columns include a ious ac o s ha could in luence he p obabili y o
u no e .
38
• u no e : The a ge a iable ha indica es whe he he employee expe ienced u no e (1)
o s ayed (0) in he company.
I is impo an o no e ha he alue in he
u no e
column e e s o he employee’s u no e s a us
in he yea ollowing he one lis ed in he yea column. Fo example, i
yea = 2022
and
u no e = 1
, i
means ha he employee le he company in 2023. This app oach is used because he goal o he model
is o o ecas u no e in he upcoming yea s based on his o ical da a up o he p esen .
This s uc u e acili a es no only he his o ical acking o each employee bu also he inpu o a ious
p edic o s in o he model. In a scena io like 2024, he objec i e is o p edic he p obabili y o u no e
o he upcoming i e yea s using he in o ma ion a ailable up o ha poin .
4.1.3 Va iable S uc u ing
To imp o e cla i y and acili a e unde s anding, he a iables we e di ided in o i e main ca ego ies.
Below a e he key a iables used in he p edic i e model, along wi h hei desc ip ions:
1. Cu en Posi ion
• S o e Type: Re e s o he di e en ypes o s o es o di isions wi hin he company.
• Role Clus e : Desc ibes he employee’s unc ion wi hin he company.
• S o e Region and Employee Region: The geog aphical loca ion o he s o e and he
employee’s esidence.
• S o e Dis ic and Employee Dis ic : The dis ic whe e he s o e and employee a e loca ed.
• Con ac Type: Indica es whe he he employee has a ixed- e m o inde ini e con ac .
• O ganiza ional Uni Change: Re e s o he numbe o changes in he o ganiza ional uni
h oughou he yea .
• Leade ship Change: The numbe o leade ship changes he employee expe ienced du ing
he yea .
• Team Exi s unde same Leade : The numbe o eam membe s unde he same leade who
le olun a ily in ha yea .
• Wo kload Type: Whe he he employee wo ks ull- ime o pa - ime.
2. Ex e nal Fac o s
• A e age Wage Va ia ion: The a ia ion in a e age wages compa ed o he p e ious yea .
• Minimum Wage Va ia ion: The a ia ion in he minimum wage compa ed o he p e ious
yea .
• In la ion Ra e: Measu es he gene al p ice a ia ion in he economy.
39
• Unemploymen Ra e: Re lec s labo ma ke condi ions.
3. Compensa ion and Pe o mance
• Theo e ical Fixed Compensa ion and Va iable Ta ge Compensa ion: Annual compensa ion
adjus ed o he mon hly minimum wage.
• Ta ge Achie emen : The pe cen age o a iable a ge achie emen .
• Change in Ta ge Achie emen : The di e ence in a ge achie emen compa ed o he
p e ious yea .
• Pe o mance: The employee’s pe o mance, e alua ed on a scale o 1 o 5.
• Po en ial: The employee’s po en ial, e alua ed on a scale o 1 o 3.
• Pe o mance Change and Po en ial Change: Pe o mance and po en ial a ia ions compa ed
o he p e ious yea .
• Ex a Compensa ion: Addi ional compensa ion o ex a hou s o s enuous wo k.
• Time Since Las P omo ion: Time since he employee’s las p omo ion.
• A e age Wo kload: The a e age weekly wo kload.
• Wo kload Va ia ion: The s anda d de ia ion o he employee’s weekly wo kload.
• Absence Hou s and Team Absence Hou s: The numbe o hou s o absence o he employee
and he eam.
• T aining Hou s: The numbe o aining hou s he employee has ecei ed.
4. Employee Cha ac e is ics
• Age: The employee’s age.
• Tenu e: The employee’s leng h o se ice in he company.
• Academic Quali ica ions: Re e s o he highes le el o educa ion a ained by he employee.
• Numbe o Dependen s: The numbe o dependen s he employee has.
• Ma i al S a us: The employee’s ma i al s a us.
5. Leade ship Pe o mance
• Sel -Assessmen by he Leade : The leade ’s sel -e alua ion o hei pe o mance.
• Team Assessmen o Leade ship: The e alua ion o leade ship made by he eam membe s.
• Leade ship Assessmen Di e ence: The di e ence be ween he leade ’s sel -assessmen and
he eam’s e alua ion.
Some a iables unde wen speci ic p ocesses o be accu a ely ep esen ed in he model. These
ans o ma ions ensu e a be e e lec ion o he unde lying dynamics. Below a e he key a iable
adjus men s:
40
• Wage Va ia ion: Va iables like
a e age wage a ia ion
and
minimum wage a ia ion
we e
calcula ed as he pe cen age a ia ion be ween he cu en yea and he p e ious yea using he
o mula:
𝑊𝑎𝑔𝑒 𝑉𝑎𝑟𝑖𝑎𝑡𝑖𝑜𝑛 = 𝐶𝑢𝑟𝑟𝑒𝑛𝑡 𝑌𝑒𝑎𝑟 𝑉𝑎𝑙𝑢𝑒−𝑃𝑟𝑒𝑣𝑖𝑜𝑢𝑠 𝑌𝑒𝑎𝑟 𝑉𝑎𝑙𝑢𝑒
𝐶𝑢𝑟𝑟𝑒𝑛𝑡 𝑌𝑒𝑎𝑟 𝑉𝑎𝑙𝑢𝑒 (4.1)
o Ad an age: Calcula ing wage a ia ions as a pe cen age be ween consecu i e yea s
allows he model o cap u e no jus absolu e changes in wages, bu also how signi ican
hose changes a e ela i e o p e ious alues. This app oach be e e lec s economic
ends, such as in la ion o wage s agna ion and helps o no malize he da a o e ime.
I highligh s he impac o wage luc ua ions, gi ing he model a be e unde s anding o
how changing economic condi ions migh in luence an employee’s decision o s ay o
lea e.
• Compensa ion No maliza ion: The a iables
heo e ical ixed compensa ion
and
a iable a ge
compensa ion
we e di ided by he minimum wage o each espec i e yea .
o Ad an age: By di iding
heo e ical ixed compensa ion
and
a iable a ge compensa ion
by he minimum wage o each yea , he model accoun s o in la ion and pu chasing
powe . This no maliza ion is c ucial o compa ing compensa ion ac oss di e en ime
pe iods since a highe nominal sala y in a la e yea migh no ep esen a ue inc ease
in alue due o in la ion. This adjus men ensu es ha he model is e alua ing sala ies
in eal e ms, making i mo e accu a e in assessing he po en ial impac o compensa ion
on employee u no e .
• Ta ge Achie emen : This ep esen s he achie emen a e o he a ge a iable compensa ion,
calcula ed as:
𝑇𝑎𝑟𝑔𝑒𝑡 𝐴𝑐ℎ𝑖𝑒𝑣𝑒𝑚𝑒𝑛𝑡 = 𝑎𝑚𝑜𝑢𝑛𝑡 (𝑖𝑛𝑐𝑒𝑛𝑡𝑖𝑣𝑒𝑠)
𝑉𝑎𝑟𝑖𝑎𝑏𝑙𝑒 𝑇𝑎𝑟𝑔𝑒𝑡 𝐶𝑜𝑚𝑝𝑒𝑛𝑠𝑎𝑡𝑖𝑜𝑛 (4.2)
o Ad an age: Calcula ing a ge achie emen as he a io o incen i es o a iable a ge
compensa ion helps quan i y how well employees a e pe o ming ela i e o hei a iable
compensa ion goals. This me ic can e eal pa e ns in employee engagemen and
41
mo i a ion. Fo ins ance, consis en unde pe o mance migh co ela e wi h a highe
likelihood o u no e , while exceeding a ge s may indica e s abili y.
• Change in Ta ge Achie emen : The a ia ion in a ge achie emen be ween consecu i e yea s,
calcula ed as:
𝐶ℎ𝑎𝑛𝑔𝑒 𝑇𝑎𝑟𝑔𝑒𝑡 𝐴𝑐ℎ𝑖𝑒𝑣𝑒𝑚𝑒𝑛𝑡 = 𝑇𝑎𝑟𝑔𝑒𝑡 𝑎𝑐ℎ𝑖𝑒𝑣𝑒𝑚𝑒𝑛𝑡 [𝑛]
𝑇𝑎𝑟𝑔𝑒𝑡 𝑎𝑐ℎ𝑖𝑒𝑣𝑒𝑚𝑒𝑛𝑡 [𝑛−1] − 1 (4.3)
o Ad an age: By calcula ing he pe cen age change in a ge achie emen be ween
consecu i e yea s, he model can ack whe he an employee’s pe o mance is imp o ing
o declining. Sudden d ops in a ge achie emen migh signal dissa is ac ion o
disengagemen , which can se e as ea ly indica o s o u no e . This dynamic iew is
mo e insigh ul han simply acking absolu e pe o mance.
• Pe o mance and Po en ial Changes: The a iables
pe o mance change
and
po en ial change
cap u e he change in pe o mance and po en ial by sub ac ing he p e ious yea ’s sco e om
he cu en yea ’s sco e:
𝑃𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒/𝑃𝑜𝑡𝑒𝑛𝑡𝑖𝑎𝑙 𝐶ℎ𝑎𝑛𝑔𝑒𝑠 = 𝐶𝑢𝑟𝑟𝑒𝑛𝑡 𝑌𝑒𝑎𝑟 𝑆𝑐𝑜𝑟𝑒 −
𝑃𝑟𝑒𝑣𝑖𝑜𝑢𝑠 𝑌𝑒𝑎𝑟 𝑆𝑐𝑜𝑟𝑒 (4.4)
o Ad an age: Using he di e ence be ween he cu en and p e ious yea ’s pe o mance
and po en ial sco es allows he model o de ec shi s in an employee's ca ee ajec o y.
A consis en decline in pe o mance o po en ial migh indica e bu nou o a lack o u u e
oppo uni ies, whe eas imp o emen could signal inc eased e en ion likelihood. The
model, he e o e, ge s a clea e sense o an employee’s momen um a he han jus a
s a ic snapsho .
• Addi ional Compensa ion: This a iable was calcula ed as he a io o ex a compensa ion o ixed
compensa ion:
𝐴𝑑𝑑𝑖𝑡𝑖𝑜𝑛𝑎𝑙 𝐶𝑜𝑚𝑝𝑒𝑛𝑠𝑎𝑡𝑖𝑜𝑛 = 𝑇𝑜𝑡𝑎𝑙 𝐸𝑥𝑡𝑟𝑎 𝐶𝑜𝑚𝑝𝑒𝑛𝑠𝑎𝑡𝑖𝑜𝑛
𝐹𝑖𝑥𝑒𝑑 𝐶𝑜𝑚𝑝𝑒𝑛𝑠𝑎𝑡𝑖𝑜𝑛 (4.5)
o Ad an age: Calcula ing he a io o ex a compensa ion o ixed compensa ion p o ides
insigh in o how much addi ional wo kload o s enuous condi ions a e being
compensa ed o . I employees a e egula ly ecei ing a high amoun o ex a
48
So, as a esume, he a iables ha a e aken in o conside a ion in he i e-yea model a e:
Table 5 - Va iables Classi ica ion o Tu no e Fi e-yea P edic ion Model
Ca ego ical Va iables
Nume ical Va iables
S o e Type
A e age Wage
Role Clus e
Minimum Wage
S o e Dis ic
In la ion Ra e
Academic Quali ica ions
Unemploymen Ra e
Con ac Type
Theo e ical Fixed Compensa ion
Wo kload Type
Pe o mance
Po en ial
Age
Tenu e
Tu no e ( a ge a iable)
The decision o use absolu e alues o minimum and medium wages, ins ead o a ia ions, is
due o he p ac icali y o long- e m scena io planning. Wi h absolu e alues, he clien can di ec ly inpu
p ojec ed wage le els wi hou needing o calcula e annual changes, simpli ying he e alua ion o a ious
sala y s a egies on u no e o e he i e-yea pe iod. This app oach p o ides an in ui i e and e icien
way o simula e and assess he impac s o di e en emune a ion s a egies, suppo ing mo e s a egic
decision-making.
The i e-yea model aims o p edic employee u no e o e a longe ime ame using
pe sonas
a he han indi idual employee da a. Like he one-yea model, i u ilizes he XGBoos algo i hm due o i s
e ec i eness wi h he e ogeneous da a, handling o missing alues and p e en ion o o e i ing. This
model o ecas s u no e as a con inuous a iable (a e age u no e a e) ac oss g ouped
pe sonas
,
le e aging XGBoos ’s abili y o handle bo h ca ego ical and con inuous a ge a iables, making i well-
sui ed o he nuanced needs o his ex ended p edic ion model.
Da a Spli ing
Fo he i e-yea model, he da a spli ing app oach is adjus ed o handle he agg ega ed na u e
o he model, whe e p edic ions a e made a he
pe sona
le el a he han he indi idual employee le el.
Since i ’s imp ac ical o p edic exac u no e a he employee le el o e i e yea s, we use agg ega ed
pe sona
da a. The da a om 2023 is ese ed o es ing, allowing he model o be alida ed using he
mos ecen da a and simula ing a ealis ic i e-yea o ecas .
49
Unlike he one-yea model, a calib a ion se is unnecessa y, as we’ e dealing wi h a con inuous
a ge a iable (a e age u no e a e) ins ead o a ca ego ical one. None heless, i ’s c ucial o ensu e
ha he aining da a e lec s he pe iod being o ecas ed, cap u ing gene al u no e ends o e ime
accu a ely.
Model Cons uc ion
The i e-yea u no e model builds on he same XGBoos amewo k as he one-yea model,
e aining many o he co e hype pa ame e s due o hei e ec i eness in handling complex,
he e ogeneous da ase s. Howe e , gi en he dis inc equi emen s o o ecas ing o e a longe ime ame
and p edic ing a con inuous a iable, ce ain adjus men s we e made o enhance he model's
pe o mance:
The i e-yea u no e model builds on he same XGBoos amewo k as he one-yea model,
e aining many o he co e hype pa ame e s due o hei e ec i eness in handling complex,
he e ogeneous da ase s. Howe e , gi en he dis inc equi emen s o o ecas ing o e a longe ime ame
and p edic ing a con inuous a iable, ce ain adjus men s we e made o enhance he model's
pe o mance:
• Numbe o T ees (n ees): Inc eased om 50 o 100. Wi h he i e-yea model o ecas ing o e a
longe pe iod, ha ing mo e ees allows o g ea e complexi y and enables he model o cap u e
in ica e pa e ns wi hin he agg ega ed da a. This dep h is pa icula ly use ul when ying o
model nuanced, long- e m ends ac oss di e se
pe sonas
, whe e mo e decision pa hs con ibu e
o highe accu acy in con inuous p edic ions.
• Maximum Dep h (max_dep h): Raised om 6 o 10. By inc easing he ee dep h, he model can
explo e deepe in e ac ions be ween a iables, which is essen ial o a model ope a ing on
pe sona
-le el agg ega ions a he han indi idual da a. This allows he model o cap u e de ailed
ela ionships wi hin he da a ha migh in luence u no e o e a i e-yea ho izon, such as how
di e en
pe sona
ai s in e ac wi h ex e nal ac o s like in la ion o wage changes.
• Lea ning Ra e (lea n_ a e): Adjus ed om 0.2 o 0.3. Wi h he longe o ecas ho izon and
con inuous a ge , a sligh ly highe lea ning a e enables as e adjus men s in each boos ing
ound, helping he model con e ge mo e quickly on meaning ul pa e ns in he da a. This
adjus men s ikes a balance be ween allowing he model o lea n e ec i ely om he da a while
a oiding o e i ing, which is c ucial gi en he model's long- e m pe spec i e.
50
• Dis ibu ion (dis ibu ion): Se o gaussian, e lec ing he con inuous na u e o he a ge a iable.
Unlike he one-yea model, which o ecas s a bina y ou come ( u no e o no u no e ), he i e-
yea model p edic s an a e age u no e a e as a con inuous a iable. The gaussian dis ibu ion
is op imal o his ype o ou pu as i models con inuous alues and helps he model gene a e
smoo he , mo e ealis ic p edic ions o e ime.
• Calib a ion: No used in his model. Calib a ion is gene ally applied o e ine p obabili y es ima es
in ca ego ical models, pa icula ly in imbalanced da ase s. Since he i e-yea model’s a ge is a
con inuous a e age u no e a e, calib a ion would no p o ide addi ional bene i . The con inuous
na u e o he p edic ions aligns na u ally wi h he model's objec i e o es ima e a e age u no e ,
making calib a ion unnecessa y.
By ailo ing hese hype pa ame e s, he i e-yea model emains obus enough o handle la ge,
agg ega ed da ase s while o e ing he p ecision equi ed o long- e m u no e o ecas s. The o he
hype pa ame e s emain aligned wi h hose o he one-yea model, as hey al eady suppo he model's
abili y o handle la ge olumes o di e se da a e ec i ely.
4.3.3 E alua ion
Bo h models we e e alua ed using Mean Absolu e E o (MAE), B ie Sco e and de ailed subg oup
analyses o compa e p edic ed and ac ual u no e . Addi ionally, he one-yea model was also assessed
using AUC-ROC, gi en i s bina y classi ica ion na u e.
AUC-ROC
The A ea Unde he Recei e Ope a ing Cha ac e is ic Cu e (AUC-ROC) was used o e alua e
he model’s abili y o dis inguish be ween u no e and non- u no e cases. The Recei e Ope a ing
Cha ac e is ic (ROC) cu e plo s he ue posi i e a e ( he p opo ion o co ec ly p edic ed u no e
cases) agains he alse posi i e a e ( he p opo ion o non- u no e cases inco ec ly classi ied as
u no e ) ac oss di e en p obabili y h esholds. This cu e illus a es he ade-o be ween sensi i i y
and speci ici y a a ious h esholds.
The AUC-ROC p o ides a single alue summa izing he model’s o e all pe o mance in
dis inguishing be ween classes. An AUC-ROC alue close o 1 indica es ha he model has a high
capaci y o co ec ly di e en ia e be ween employees who lea e and hose who s ay, while a alue close
o 0.5 indica es pe o mance no be e han andom guessing.
51
Mean Absolu e E o (MAE)
The MAE was used o measu e he a e age absolu e e o s be ween he p edic ed u no e and
he ac ual alues. This me ic p o ides insigh in o how close he model's p edic ions a e o he eal
u no e a es. The lowe he MAE, he be e he model’s pe o mance, as i means ha he p edic ed
alues a e close o he ac ual ou comes.
𝑀𝐴𝐸 = 1
𝑛∑|
𝑛
𝑖=1 ŷ𝑖-𝑦𝑖| (4.7)
The MAE was used o measu e he a e age absolu e e o s be ween he p edic ed u no e
p obabili ies and he ac ual u no e alues. Al hough MAE is ypically applied o con inuous a iables, i
is also sui able in his con ex o e alua e bo h models — including he one-yea p edic ion model, whe e
he a ge a iable is bina y (0 o 1). This is because, e en hough he ac ual alues a e bina y, he model
ou pu s p obabili ies, ep esen ing he likelihood o each employee lea ing. I he model was pe ec , he
p edic ed p obabili ies would always be ei he 0 o 1, ma ching he ac ual ou comes. By using MAE, we
can assess how close he p edic ed p obabili ies a e o he co ec bina y alues, p o iding aluable
insigh in o he model’s o e all pe o mance.
B ie Sco e
The B ie Sco e was used o e alua e he accu acy o he p obabilis ic p edic ions p oduced by
he models. I measu es he mean squa ed di e ence be ween he p edic ed p obabili ies and he ac ual
ou comes. As such, i is pa icula ly sui able o assessing he calib a ion o classi ica ion models whe e
he ou pu is a p obabili y a he han a disc e e class.
𝐵𝑟𝑖𝑒𝑟 𝑆𝑐𝑜𝑟𝑒 = 1
𝑛∑(ŷ𝑖− y𝑖)2
𝑛
𝑖=1 (4.8)
In he one-yea model, whe e he a ge a iable is bina y (0 o 1), he B ie Sco e cap u es how
close he p edic ed p obabili ies a e o he ac ual bina y ou comes. A lowe B ie Sco e indica es be e
calib a ion and highe con idence in he p edic ed p obabili ies. While AUC-ROC ocuses on he model’s
anking abili y, he B ie Sco e o e s complemen a y in o ma ion by assessing whe he he p edic ed
p obabili y alues a e well-calib a ed — o example, whe he a 70% p edic ed p obabili y o u no e
co esponds o app oxima ely 70% o ac ual u no e in simila cases.
52
In he i e-yea model, whe e u no e is exp essed as a con inuous a e a he g oup le el, he
B ie Sco e also emains applicable. I con inues o measu e he squa ed e o be ween p edic ed and
obse ed a es, unc ioning as a calib a ion me ic e en in he con ex o agg ega ed p edic ions.
Subg oup analysis
In addi ion o he o e all me ics, he model was analyzed in de ail by ocusing on speci ic
ca ego ical a iables. This analysis compa ed he p edic ed u no e and he ac ual u no e wi hin
speci ic subg oups o ensu e ha he model cap u es all nuances ac oss di e en g oups o employees.
Fo ins ance, while he o al p edic ed u no e migh align well wi h he ac ual o al u no e ,
signi ican di e ences could occu wi hin speci ic subg oups such as di e en ole clus e s. The goal is
no only o p edic he o e all u no e bu also o assess he accu acy wi hin each ca ego y, ensu ing
ha he model pe o ms well ac oss di e se employee unc ions.
This de ailed analysis by subg oups (such as ole clus e s,
pe sonas
and dis ic s) is also essen ial
o eed he wo k o ce op imiza ion model, which elies on accu a e u no e p edic ions ac oss hese
di e en segmen s o suppo s a egic wo k o ce planning decisions.
4.4 Final Rema ks
This chap e p esen ed he me hodology de eloped o p edic olun a y employee u no e ac oss
wo ime ho izons: a sho - e m indi idual-le el model (one-yea ) and a long- e m agg ega e-le el model
( i e-yea ). The i s model le e ages de ailed employee-le el da a o es ima e he p obabili y o u no e
in he ollowing yea , o e ing g anula , ope a ional insigh s. The second model, based on agg ega ed
pe sonas
ac oss s o e ypes, ole clus e s and dis ic s, suppo s s a egic planning by o ecas ing
expec ed u no e olumes o e a longe ho izon.
Bo h models we e implemen ed using he XGBoos algo i hm, selec ed o i s abili y o handle
he e ogeneous and imbalanced da a while main aining scalabili y and in e p e abili y — essen ial ea u es
in a complex, da a- ich en i onmen such as e ail. The me hodology inco po a ed class balancing
s a egies, ca e ul a iable design and obus e alua ion me ics — including MAE, B ie Sco e and AUC-
ROC — complemen ed by subg oup analyses o ensu e pe o mance ac oss di e se wo k o ce segmen s.
Toge he , hese me hodological choices aim o suppo in o med, da a-d i en decision-making a
bo h ac ical and s a egic le els, aligning p edic i e powe wi h eal-wo ld wo k o ce planning needs. The
nex chap e builds upon his ounda ion by p esen ing he p edic i e pe o mance o each model and
53
iden i ying he mos in luen ial u no e d i e s, p o iding ac ionable insigh s o imp o e employee
e en ion and o ganiza ional s abili y.
54
5. RESULTS ANALYSIS AND DISCUSSION
This chap e p esen s he esul s o he p edic i e models de eloped o es ima e olun a y
u no e o e one- and i e-yea ho izons. I begins by e alua ing he pe o mance o he one-yea model,
which p edic s indi idual-le el u no e based on de ailed employee ea u es. Subsequen ly, he chap e
discusses he esul s o he i e-yea model, which o ecas s agg ega e u no e a a
pe sona
-le el ac oss
di e en o ganiza ional segmen s. Fo bo h models, he analysis conside s p edic i e accu acy, e o
dis ibu ion ac oss subg oups and he impo ance o key ea u es d i ing u no e . These esul s a e hen
in e p e ed in ligh o he p ojec 's objec i es, p o iding insigh s in o he model's s a egic applicabili y in
wo k o ce planning.
5.1 One-yea Model
5.1.1 Assessmen o P edic i e Pe o mance
The one-yea u no e p edic ion model was e alua ed using ou key me ics: AUC-ROC, Mean
Absolu e E o (MAE), BIAS and B ie Sco e. Each o hese me ics cap u es di e en aspec s o model
pe o mance, om i s abili y o ank employees by u no e isk o how well he p edic ed p obabili ies
align wi h ac ual u no e a es. Table 6 summa izes he esul s.
Table 6 - O e all Pe o mance Me ics o he One-Yea Tu no e P edic ion Model
AUC-ROC
BIAS
MAE
B ie Sco e
93 %
1.17 pp
10.36 pp
0.046
An AUC-ROC o 93% sugges s ha he model e ec i ely anks employees based on hei likelihood
o lea ing. This means ha indi iduals wi h highe p edic ed u no e p obabili ies a e indeed mo e likely
o lea e. Howe e , a high AUC does no gua an ee well-calib a ed p obabili ies, meaning ha while he
model can o de employees co ec ly, he ac ual p obabili y alues may s ill equi e e inemen .
The MAE o 10.36 pe cen age poin s (pp) indica es ha , on a e age, p edic ed u no e
p obabili ies de ia e om ac ual ou comes by abou 10 pp. Conside ing ha he o e all u no e a e is
11.9%, his e o ep esen s a subs an ial p opo ion o he expec ed alue and is he e o e conside ed
high in his con ex . While some deg ee o e o is o be expec ed in p obabilis ic models, such a de ia ion
limi s he model’s use ulness o accu a e planning a he indi idual le el and calls o ca e ul agg ega ion
and alida ion a he g oup le el o ensu e eliable wo k o ce planning ou pu s.
55
To complemen his, he B ie Sco e o he model is 0.046, indica ing ha he p edic ed
p obabili ies a e easonably well-calib a ed. This suppo s he model’s sui abili y o agg ega ed
o ecas ing, e en i indi idual-le el p edic ions ca y highe e o ma gins.
Wi h his in mind, he o e all u no e p edic ion o he es se can be isualized in Figu e 14,
which compa es he model’s o ecas wi h he ac ual u no e a e o 2023.
The model p edic ed a u no e a e o 13.0%, compa ed o he ac ual u no e o 11.9%, esul ing
in a small o e es ima ion o 1.1 pp (Figu e 14). This sugges s ha he model success ully cap u es he
o e all end in wo k o ce a i ion, making i a aluable ool o wo k o ce planning. Howe e , while he
p edic ion is accu a e a a global le el, i is essen ial o u he in es iga e how well he model pe o ms
ac oss di e en employee segmen s.
Figu e 14 - Ac ual Tu no e s Tu no e P edic ion (One-yea model)
While he global u no e p edic ion p o ides an o e all assessmen o he model’s accu acy, i
is essen ial o analyze i s pe o mance a he segmen le el. Since wo k o ce planning decisions will be
based on agg ega ed u no e es ima es pe dis ic , s o e ype, ole clus e and
pe sona
, e alua ing he
model’s e ec i eness wi hin hese key g oupings is c ucial.
To assess his, he model's e o s we e analyzed a wo le els o agg ega ion:
• Dis ic x S o e Type x Role Clus e : A co e le el used in wo k o ce planning.
• Dis ic x S o e Type x Role Clus e x
Pe sona
: The mos g anula le el, which di ec ly eeds in o
he wo k o ce op imiza ion model.
56
A he dis ic x s o e ype x ole clus e le el, he model shows an imp o emen ac oss all me ics
compa ed o he indi idual-le el e alua ion (E o ! No a alid bookma k sel - e e ence.). The MAE
is educed by mo e han hal , he B ie Sco e dec eases en old and he BIAS inc eases mode a ely om
1.17 pp o 1.93 pp. This indica es ha agg ega ing p edic ions a his le el helps smoo h ou indi idual
a iabili y, esul ing in mo e s able and be e -calib a ed o ecas s. These indings suppo he use o ole
clus e as a sui able le el o g anula i y o s a egic wo k o ce planning.
A he dis ic x s o e ype x ole clus e ×
pe sona
le el, howe e , he bene i s o agg ega ion a e less
clea . While he MAE dec eases o 9.17 pp, he BIAS inc eases o 5.31 pp, e lec ing a s onge di ec ional
e o . The B ie Sco e imp o es o 0.021, indica ing be e calib a ion han a he indi idual le el, bu
emains subs an ially highe han he 0.004 obse ed a he ole clus e le el. This sugges s ha he
model deli e s less consis en and eliable p edic ions when applied a he
pe sona
le el.
Table 7 – Tu no e P edic ion Model Pe o mance a Agg ega ed Segmen Le els (One-yea model)
Le el
BIAS
MAE
B ie Sco e
Dis ic x S o e Type x Role Clus e
1.93 pp
4.42 pp
0.004
Dis ic x S o e Type x Role Clus e x
Pe sona
5.31 pp
9.17 pp
0.021
Gi en ha he op imiza ion model ope a es on
pe sona
-le el o ecas s, i becomes impo an o
assess whe he he obse ed disc epancies s em om model limi a ions o om speci ic cha ac e is ics
associa ed wi h each employee segmen . To suppo his analysis, he ollowing sec ion e alua es model
pe o mance ac oss a se o ele an segmen s, including key
pe sona
-de ining a iables — such as
academic quali ica ions, con ac ela ionship, wo kload, pe o mance, po en ial, enu e and age — as
well as o ganiza ional segmen s like s o e ype and ole clus e . This analysis aims o iden i y po en ial
sou ces o e o and assess whe he he cu en segmen a ion s a egy su icien ly cap u es employee
beha io .
Figu e 15 compa es he p edic ed u no e a es gene a ed by he model wi h he ac ual u no e
obse ed in 2023 ac oss he op 10 dis ic s, which oge he ep esen app oxima ely 87.5% o he
wo k o ce.
In hese co e dis ic s, p edic ion e o s ange om 1 o 3 pe cen age poin s — well below he
lowes MAE eco ded ac oss p e iously e alua ed segmen a ions (4.42 pp a he ole clus e le el). This
indica es ha , e en hough dis ic -le el p edic ions we e no he di ec a ge o model calib a ion, he
esul s a e highly accu a e and s able ac oss key egions.
57
The e o is consis en ly posi i e, wi h he model o e es ima ing u no e ac oss all dis ic s. Fo
ins ance, he de ia ions each 3 pp in Po o and A ei o, while in dis ic s like Lisboa, Fa o, B aga and
Lei ia he gap is 1 o 2 pp. In Ilha da Madei a, he model shows an especially small de ia ion (1 pp),
aligning closely wi h he obse ed u no e in ha egion.
This sys ema ic o e es ima ion mi o s he di ec ional bias p e iously iden i ied (BIAS = 1.93 pp),
bu he low magni ude o e o s ac oss dis ic s ein o ces he model’s obus ness o geog aphic
segmen a ion. These esul s suppo he eliabili y o dis ic -le el p edic ions as an inpu o egionally
a ge ed wo k o ce planning and u no e mi iga ion s a egies.
Figu e 15 - Ac ual Tu no e s Tu no e P edic ion by Top 10 Dis ic (One-yea model)
As seen in he Figu e 16, he model once again pe o ms be e o he segmen s wi h highe
employee ep esen a ion. P edic ions a e closes o hype ma ke s and supe ma ke s, whe e mos
employees a e concen a ed (app oxima ely 80%). The di e ence be ween p edic ed and ac ual u no e
is only 1 pp o hype ma ke s and 2 pp o supe ma ke s, while o con enience s o es, he gap inc eases
o 3 pp.
As obse ed in o he analyses, he model ends o sys ema ically o e es ima e u no e , which could
indica e a sligh calib a ion bias bu also o e s a consis en pa e n ha can be add essed h ough pos -
64
5.1.2 D i e s o Tu no e P edic ion
Unde s anding which ea u es con ibu e mos o he model's p edic ions is essen ial no only o
in e p e ing esul s, bu also o guiding s a egic HR decision-making. Table 8 p esen s he en mos
impo an a iables in he one-yea u no e p edic ion model, based on hei scaled impo ance ( ela i e
o he mos impac ul a iable) and pe cen age con ibu ion o he model’s o e all p edic i e powe .
Table 8 - Top 10 Mos Impo an Va iables (One-yea Model)
Va iable
Scaled Impo ance (%)
Pe cen age (%)
Theo e ical Fixed Compensa ion
100.00
58.33
Senio i y
25.46
14.85
Age
9.52
5.55
Unemploymen Ra e
3.71
2.16
T aining Hou s
3.15
1.84
A e age Wage Va ia ion
3.11
1.81
Wo kload Type
2.65
1.54
Dis ic
2.50
1.46
Team Absence Hou s
2.08
1.22
Absence Hou s
1.86
1.09
The en mos impo an a iables, shown in Table 8, a e anked acco ding o hei scaled and
pe cen age impo ance wi hin he XGBoos amewo k. Collec i ely, hese en a iables accoun o 89.9%
o he model’s o al p edic i e weigh , indica ing ha he model’s decisions a e highly concen a ed in a
small g oup o ea u es.
Among hese, he op h ee a iables — Theo e ical Fixed Compensa ion, Senio i y and Age — a e
pa icula ly dominan , oge he ep esen ing 78.73% o he o al impo ance. This highligh s he model’s
s ong eliance on hese ac o s when es ima ing u no e p obabili y. The Theo e ical Fixed
Compensa ion alone con ibu es 58.33% o he o al weigh , clea ly s anding ou as he mos in luen ial
ea u e. Fo con ex , in ela i e e ms (scaled impo ance), Senio i y holds 25.46% o he impo ance o
Compensa ion, while Age ep esen s 9.52%. These di e ences illus a e how cen al compensa ion is o
he model’s p edic ions, while senio i y and age, hough s ill ele an , play mo e complemen a y oles.
In con as , he emaining se en a iables each con ibu e less han 3% o he model’s decisions,
ein o cing he idea ha he model ocused mos o i s p edic i e powe on a na ow se o ea u es —
65
pa icula ly on emune a ion. This le el o concen a ion is aluable o in e p e a ion, as i p o ides clea
insigh in o which dimensions a e mos s ongly associa ed wi h u no e isk in he e ail wo k o ce
con ex .
Beyond iden i ying he indi idual con ibu ion o each ea u e, i is also possible o d aw insigh s
by g ouping a iables in o hema ic clus e s. These clus e s we e p e iously de ined based on he na u e
o he a iables: Compensa ion and Pe o mance, Employee Cha ac e is ics, Cu en Posi ion, Ex e nal
Fac o s and Leade ship Pe o mance. F om his pe spec i e, Compensa ion and Pe o mance eme ges
as he mos dominan g oup, con ibu ing wi h o e 60% o he o al model impo ance. This is la gely
due o he weigh o he Theo e ical Fixed Compensa ion, bu also suppo ed by a iables such as T aining
Hou s and A e age Wage Va ia ion. Employee Cha ac e is ics, ep esen ed mainly by Senio i y and Age,
join ly accoun o o e 20% o he o al impo ance, highligh ing he ole o employee enu e and li ecycle
in u no e isk. Va iables ela ed o he Cu en Posi ion, such as Wo kload Type, Team Absence Hou s,
Absence Hou s and Dis ic , ep esen a smalle po ion o he impo ance — collec i ely unde 6%.
Simila ly, Ex e nal Fac o s such as Unemploymen Ra e and A e age Wage Va ia ion ha e mode a e
ele ance (a ound 4%), sugges ing ha b oade labo ma ke condi ions, while no negligible, a e
seconda y o in e nal cha ac e is ics. This b eakdown ein o ces he idea ha compensa ion and
indi idual employee cha ac e is ics a e he p ima y d i e s in p edic ing u no e in he sho e m,
whe eas con ex ual o o ganiza ional a iables play a suppo ing ole.
5.2 Fi e-yea model
5.2.1 Assessmen o P edic i e Pe o mance
Since he i e-yea model p edic s u no e as a con inuous a iable — based on a e age u no e
a es wi hin each dis ic × s o e ype × ole clus e × pe sona — adi ional classi ica ion me ics such as
AUC-ROC a e no applicable, as he e is no bina y g ound u h o ank o classi y. Ins ead, model
pe o mance is e alua ed using eg ession-o ien ed me ics: Mean Absolu e E o (MAE), BIAS, and B ie
Sco e. These me ics p o ide complemen a y insigh s in o he model’s beha io , cap u ing aspec s such
as a e age de ia ion, di ec ional bias, and calib a ion. The esul s a e p esen ed in Table 9.
Table 9 - O e all Pe o mance Me ics o he Fi e-Yea Tu no e P edic ion Model
BIAS
MAE
B ie Sco e
- 1.2 pp
15.97 pp
0.062
66
The MAE o 15.97 pp indica es ha , on a e age, he p edic ed u no e a es de ia e om he
ac ual alues by app oxima ely 16 pe cen age poin s. While his e o is highe han he 10.36 pp
obse ed in he one-yea model, i is expec ed gi en he inc eased unce ain y associa ed wi h long- e m
o ecas s and he lack o indi idual-le el da a o u u e employees.
The B ie Sco e o 0.062 e lec s he a e age squa ed di e ence be ween he p edic ed
p obabili ies and ac ual ou comes. Despi e he model ope a ing a an agg ega ed le el and o e a longe
ime ame, his ela i ely low sco e sugges s ha he model main ains a easonable deg ee o calib a ion
— meaning ha , in mos segmen s, he p edic ed u no e a es closely e lec ac ual g oup beha io .
The BIAS o -1.2 pp e eals a sligh unde es ima ion end, wi h he model p edic ing, on a e age,
lowe u no e han wha was obse ed. This con as s wi h he one-yea model, which showed a posi i e
bias (o e es ima ion). This change in di ec ion may be explained by s uc u al di e ences be ween he
wo models, pa icula ly he use o agg ega ed g oup-le el inpu s and he inc eased complexi y o long-
e m p edic ion. Addi ionally, such di ec ional e o s could s em om unce ain y p opaga ion,
unobse ed coho e ec s o a ia ion in how
pe sonas
e ol e o e ime.
Toge he , hese esul s con i m ha he i e-yea model is capable o p oducing meaning ul long-
e m o ecas s, despi e wo king unde mo e cons ained condi ions. Howe e , due o i s highe e o and
unde es ima ion endency, i is impo an ha i s ou pu s a e in e p e ed wi h cau ion and supplemen ed
wi h scena io es ing — pa icula ly when being used o in o m s a egic decisions o e mul i-yea ho izons.
Figu e 25 - Ac ual Tu no e s Tu no e P edic ion (Fi e-yea model)
Figu e 25 displays he compa ison be ween he model’s p edic ed u no e and he ac ual
u no e obse ed in 2023, based on agg ega ed alues ac oss all segmen s in he es da ase . The i e-
yea model p edic ed an a e age u no e a e o 13.7%, while he ac ual u no e was 11.9%, esul ing
in a 1.8 pp o e es ima ion a he global le el.
67
Despi e he highe p edic ed u no e a e (13.7%) compa ed o he ac ual u no e obse ed in
2023 (11.9%), he model p esen s a nega i e BIAS o –1.2 pp. While his may ini ially appea
con adic o y, i is impo an o no e ha BIAS e lec s he a e age di ec ion o he p edic ion e o s ac oss
all segmen s (i.e., dis ic × s o e ype × ole clus e ×
pe sona
), a he han he o e all di e ence be ween
agg ega ed alues. In his case, he nega i e BIAS indica es ha he model ends o unde es ima e
u no e in mos segmen s, e en hough a smalle numbe o segmen s wi h la ge o e es ima ions
con ibu e o an o e all p edic ed u no e ha exceeds he ac ual igu e. This nuance highligh s he need
o in e p e BIAS and agg ega e esul s oge he , as hey cap u e di e en aspec s o model pe o mance.
While he global me ics p esen ed ea lie o e a i s assessmen o he model’s pe o mance,
hey a e calcula ed o e a ine-g ained segmen a ion ha includes all combina ions o he ca ego ical
a iables — no he
pe sonas
de ined by he company. In ac , hese combina ions include all possible
g oupings ac oss dis ic , s o e ype, ole clus e and all
pe sona
-de ining a iables indi idually, a he
han he speci ic
pe sona
g oupings used in wo k o ce planning. Fo his eason, i is essen ial o explici ly
assess model pe o mance a wo addi ional le els: i s , a he dis ic × s o e ype × ole clus e le el, o
analyze a less g anula bu ope a ionally ele an segmen ; and second, a he
pe sona
le el, which is he
exac agg ega ion used o eed he op imiza ion model. This dis inc ion is c i ical, as he me ics a he
pe sona
le el — whe e a ibu es like age o pe o mance a e g ouped in o p ede ined bands — can di e
subs an ially om hose calcula ed on ully disagg ega ed da a. The ollowing esul s ocus on e alua ing
he model’s beha io wi hin hese wo key le els o agg ega ion.
The pe o mance esul s in Table 10 e eal ha model accu acy imp o es signi ican ly when
p edic ions a e agg ega ed a he ole clus e le el, wi h Mean Absolu e E o (MAE) d opping om 15.97
pp a he global le el o jus 6.63 pp. This subs an ial educ ion highligh s he alue o agg ega ion in
smoo hing indi idual-le el a ia ion and p oducing mo e s able es ima es. The B ie Sco e also dec eases
sha ply o 0.008, e lec ing imp o ed calib a ion o he p edic ed p obabili ies a his less g anula le el.
Mo eo e , he BIAS shi s om –1.2 pp in he o e all model o +0.94 pp when p edic ions a e
agg ega ed a he dis ic × s o e ype × ole clus e le el, sugges ing ha while he model ends o
unde es ima e u no e globally, i sligh ly o e es ima es u no e a his ope a ionally ele an
agg ega ion. This change in di ec ion e lec s how agg ega ion can balance ou localized unde es ima ions
and p o ide mo e neu al p edic ions, especially a le els used o high-le el wo k o ce planning.
A he
pe sona
le el — he ac ual g anula i y used o eed he op imiza ion model — he me ics
a e less a o able. The MAE inc eases o 14.09 pp and he B ie Sco e ises o 0.042, indica ing ha he
model s uggles mo e o main ain accu acy and calib a ion when a ge ing speci ic
pe sona
g oups.
68
Addi ionally, he BIAS g ows o 4.18 pp, poin ing o a mo e p onounced and consis en o e es ima ion
ac oss
pe sonas
.
This deg ada ion in pe o mance is expec ed, no because he model combines mul iple a iables
— as seen in he one-yea model, which handles e en mo e de ailed da a — bu a he due o he way
pe sonas
a e cons uc ed. These
pe sonas
a e p ede ined by he clien using speci ic h esholds o
nume ical a iables (e.g., g ouping age abo e o below 35), selec ed ca ego ical combina ions and a ixed
se o a ibu es. While his app oach is ope a ionally meaning ul and essen ial o wo k o ce planning, i
in oduces a i icial segmen a ion ha may no align wi h he na u al pa e ns p esen in he da a. As a
esul , some p edic i e accu acy is los a his le el.
Table 10 - Tu no e P edic ion Model Pe o mance a Agg ega ed Segmen Le els (Fi e-yea model)
Le el
BIAS
MAE
B ie Sco e
Dis ic x S o e Type x Role Clus e
0.94 pp
6.63 pp
0.008
Dis ic x S o e Type x Role Clus e x
Pe sona
4.18 pp
14.09 pp
0.042
To gain a mo e comp ehensi e unde s anding o he model’s beha io , i is he e o e impo an
o analyze p edic ion accu acy ac oss all a iables ha de ine he o ecas g anula i y. This includes no
only he ope a ional segmen a ion a iables — dis ic , s o e ype and ole clus e — bu also he
cha ac e is ics used o cons uc he
pe sonas
. These
pe sona
-de ining a iables (e.g., age, enu e,
con ac ype, wo kload, academic quali ica ions, pe o mance and po en ial) we e selec ed by he clien
and g ouped in o p ede ined ca ego ies. As such, analyzing model pe o mance along each o hese
dimensions p o ides aluable insigh s in o which speci ic segmen s a e con ibu ing o e o and whe he
ce ain combina ions equi e u he e inemen o a ge ed scena io es ing.
Figu e 26 compa es he p edic ed and ac ual u no e a es o he op 10 dis ic s in 2023,
which collec i ely ep esen o e 87.5% o he wo k o ce. In hese dis ic s, he p edic ion e o s ange
om 1 o 5 pp. Despi e being calcula ed a he dis ic le el — a highe le el o agg ega ion han he
model’s ope a ional uni — he de ia ions emain ela i ely modes . The a e age absolu e e o ac oss
hese 10 dis ic s is app oxima ely 2.4 pp, which is subs an ially lowe han he global MAE o 15.97 pp
(Table 9) and e en below he 6.63 pp MAE obse ed a he dis ic × s o e ype × ole clus e le el (Table
10).
While dis ic s a e no he di ec p edic ion uni o he model, he esul s shown he e e lec an
agg ega ion o p edic ions made a he mo e g anula le el (dis ic × s o e ype × ole clus e ×
pe sona
).
The e o e, compa ing his a e age e o o he MAE o he dis ic × s o e ype × ole clus e le el is
69
app op ia e and ein o ces he model’s s ong accu acy when esul s a e agg ega ed geog aphically —
pa icula ly in a eas wi h highe employee concen a ion.
Despi e a consis en endency o o e es ima e u no e — echoing he posi i e gap obse ed a
he global le el — only wo dis ic s, Po o and Ilha da Madei a, exhibi de ia ions o 5 pp. These ou lie s
may indica e egional speci ici ies no ully cap u ed by he model and highligh he alue o
complemen ing p edic i e ou pu s wi h con ex ual insigh s. S ill, he o e all low e o ac oss dis ic s
suppo s he model’s obus ness o geog aphically a ge ed o ecas ing and s a egic wo k o ce planning.
In e es ingly, while he i e-yea model shows mode a e de ia ions in key dis ic s like Po o (5
pp) and Lisboa (2 pp), he one-yea model achie es highe accu acy in hese same egions — wi h e o s
o only 1 pp in Lisboa and 3 pp in Po o (Figu e 15). Toge he , hese wo dis ic s accoun o
app oxima ely 43.5% o he en i e da ase , making hei accu a e p edic ion pa icula ly ele an . This
con as highligh s how sho e - e m models may be mo e p ecise o high-densi y a eas in he nea
u u e, whe eas he i e-yea model, despi e i s b oade unce ain y, p o ides aluable s a egic insigh
o long- e m planning.
Figu e 26 - Ac ual Tu no e s Tu no e P edic ion by Top 10 Dis ic (Fi e-yea model)
The model’s pe o mance by s o e ype — con enience, supe ma ke and hype ma ke — e eals
a high le el o accu acy, wi h p edic ion e o s anging be ween 1 and 2 pe cen age poin s and an a e age
absolu e de ia ion o 1.7 pp. Al hough s o e ype ep esen s a highe le el o agg ega ion han he model's
ope a ional uni , his esul is s ill in o ma i e, as u no e planning o en conside s di e ences ac oss
e ail o ma s.
No ably, he a e age e o ac oss s o e ypes is subs an ially lowe han he global MAE o 15.97
pp (Table 9) and also lowe han he MAE o 6.63 pp obse ed a he dis ic × s o e ype × ole clus e
le el (Table 10). Howe e , his di e ence should be in e p e ed wi h cau ion, as highe le els o
agg ega ion na u ally educe a iabili y and, consequen ly, e o me ics. S ill, he abili y o he model o
70
app oxima e ac ual u no e e en a his agg ega ed le el ein o ces i s consis ency and adap abili y
ac oss o ganiza ional laye s.
As illus a ed in Figu e 27, he model sligh ly o e es ima es u no e in all h ee s o e o ma s,
wi h de ia ions o 1 pp in supe ma ke s and con enience s o es and 2 pp in hype ma ke s. The small
magni ude o hese di e ences suppo s he model’s obus ness o s o e-le el o ecas ing, ein o cing i s
alue o s a egic wo k o ce planning ac oss di e en e ail o ma s. No ably, he i e-yea model (Figu e
16) achie es e y simila accu acy le els o he one-yea model o supe ma ke s and hype ma ke s —
he wo dominan s o e o ma s — despi e he added complexi y o long- e m o ecas ing. This ein o ces
con idence in i s applicabili y ac oss key ope a ional segmen s.
Figu e 27 - Ac ual Tu no e s Tu no e P edic ion by S o e Type (Fi e-yea model)
Figu e 28 compa es he p edic ed and ac ual u no e a es o he op 10 ole clus e s in 2023,
which oge he accoun o he majo i y o he wo k o ce. Ac oss hese clus e s, p edic ion e o s ange
om 1 o 6 pe cen age poin s, wi h an a e age absolu e e o o app oxima ely 2.4 pp. This is subs an ially
lowe han he 6.63 pp MAE obse ed a he same le el o agg ega ion — dis ic × s o e ype × ole
clus e — as epo ed in Table 10. This sugges s ha , when analyzed by ole clus e , he model
demons a es s ong p edic i e accu acy, pa icula ly in he mos ep esen a i e unc ions wi hin he
o ganiza ion.
The model consis en ly o e es ima es u no e ac oss all clus e s, in line wi h he global
p edic ion pa e n. Howe e , one unc ion — Bake y/Pas y Coun e Ope a o — s ands ou wi h a
de ia ion o 6 pp, indica ing a po en ial a ea o e inemen . S ill, he o e all low magni ude o e o s
ein o ces he model’s sui abili y o wo k o ce planning when segmen ing by ole clus e .
71
These indings a e consis en wi h he one-yea model (Figu e 17), which also demons a ed
s ong alignmen ac oss co e ole clus e s, wi h simila ly small and s able de ia ions. This empo al
consis ency u he suppo s he model’s obus ness when segmen ing u no e by unc ion.
Figu e 28 - Ac ual Tu no e s Tu no e P edic ion by Role Clus e (Fi e-yea model)
When analyzing u no e by academic quali ica ions, he model success ully cap u es he gene al
end ac oss educa ional le els. As shown in Figu e 29, i sligh ly unde es ima es u no e o employees
wi h highe educa ion (by 2 pp), pe ec ly aligns p edic ions o seconda y educa ion and sligh ly
o e es ima es o hose wi h p ima y educa ion (by 3 pp). Despi e hese modes de ia ions, he model
e lec s he expec ed beha io al pa e n — employees wi h highe quali ica ions end o ha e g ea e
u no e . The ela i ely small e o s ein o ce he model’s abili y o in e nalize meaning ul ela ionships
be ween educa ional backg ound and u no e , suppo ing i s applicabili y ac oss di e se employee
p o iles.
These esul s a e consis en wi h hose o he one-yea model (Figu e 18), which also accu a ely
cap u ed he u no e g adien ac oss educa ion le els wi h simila ly low e o ma gins. This ein o ces
he model’s s abili y in e lec ing educa ion- ela ed u no e ends o e di e en o ecas ing ho izons.
72
Figu e 29 - Ac ual Tu no e s Tu no e P edic ion by Academic Quali ica ions (Fi e-yea model)
The model clea ly dis inguishes be ween he wo con ac ypes, cap u ing he s uc u al
di e ence in u no e beha io be ween ixed- e m and pe manen employees. As shown in Figu e 30,
ixed- e m con ac s exhibi signi ican ly highe ac ual u no e (24%) compa ed o pe manen con ac s
(8%) — a end ha he model success ully eplica es. While he model sligh ly unde es ima es u no e
o ixed- e m employees (22% p edic ed), i o e es ima es i o pe manen employees (10% p edic ed),
esul ing in a 2 pp di ec ional gap in each g oup. These de ia ions a e modes and ein o ce he model’s
abili y o in e nalize b oad beha io al dis inc ions ac oss con ac ypes. Gi en he s ong co ela ion
be ween con ac ela ionship and employee u no e , hese esul s alida e he inclusion o his a iable
as a key d i e in
pe sona
segmen a ion and long- e m o ecas ing. No ably, he de ia ions obse ed a e
mo e balanced han hose o he one-yea model (Figu e 19), which exhibi ed a la ge o e es ima ion o
ixed- e m con ac s. This ein o ces he i e-yea model’s abili y o gene alize ac oss segmen s wi h
a ying ep esen a ion, despi e wo king wi h mo e agg ega ed inpu s.
Figu e 30 - Ac ual Tu no e s Tu no e P edic ion by Con ac Rela ionship (Fi e-yea model)
73
Figu e 31 illus a es he model’s pe o mance ac oss di e en wo kload ypes — pa - ime and
ull- ime con ac s. The model e ec i ely cap u es he subs an ial gap in u no e beha io be ween he
wo g oups, co ec ly p edic ing a signi ican ly highe u no e a e among pa - ime employees. Fo his
g oup, i sligh ly unde es ima es u no e by 1 pp (22% p edic ed s. 23% ac ual), while o ull- ime
employees, i o e es ima es he a e by 3 pp (10% s. 7%). O e all, he model aligns closely wi h he one-
yea model in his segmen a ion (Figu e 20), con i ming i s abili y o di e en ia e u no e le els by
wo kload ype and p o ide eliable es ima es o planning pu poses.
These esul s a e aligned wi h hose o he one-yea model (Figu e 20), which also accu a ely
cap u ed he beha io al gap be ween pa - ime and ull- ime employees, albei wi h sligh ly highe
de ia ions o he pa - ime g oup. This consis ency ac oss ime ho izons s eng hens he con idence in
he model’s abili y o ep esen con ac - ela ed u no e dynamics.
Figu e 31 - Ac ual Tu no e s Tu no e P edic ion by Wo kload Type (Fi e-yea model)
Figu e 32 compa es he p edic ed and ac ual u no e a es ac oss age g oups. The model
accu a ely cap u es he gene al downwa d end in u no e as age inc eases, wi h p edic ion e o s
anging be ween 1 and 2 pe cen age poin s ac oss all segmen s. This consis en alignmen ac oss age
b acke s sugges s ha he model e ec i ely di e en ia es employee beha io by age, ein o cing i s
eliabili y o wo k o ce planning ac oss gene a ional p o iles. No ably, hese esul s closely mi o hose
ob ained wi h he one-yea model (Figu e 21), con i ming he model’s abili y o eplica e known beha io al
pa e ns e en in a long- e m o ecas ing con ex .
80
O e all, he esul s demons a e he applicabili y o p edic i e analy ics o wo k o ce planning,
o e ing a solid ounda ion o bo h sho - e m ac ical decisions and long- e m s a egic planning. Beyond
s a is ical accu acy, hese models enable HR eams o p oac i ely an icipa e u no e dynamics and be e
align wo k o ce supply wi h o ganiza ional needs.
The nex chap e builds upon hese indings h ough a c i ical e lec ion, assessing he
me hodological, s uc u al and s a egic implica ions o he models and iden i ying oppo uni ies o
u he de elopmen .
81
6. CONCLUSION
This chap e concludes he s udy by e isi ing he p ojec ’s o e a ching objec i es and
summa izing i s key con ibu ions. The cen al aim was o de elop p edic i e models capable o es ima ing
olun a y employee u no e in he e ail sec o o e bo h sho - and long- e m ho izons, spanning up o
i e yea s. By applying people analy ics echniques and le e aging decision ee-based algo i hms —
pa icula ly XGBoos — he s udy demons a ed ha u no e can be e ec i ely o ecas ed using a
combina ion o in insic and ex insic employee cha ac e is ics. Beyond p edic i e accu acy, he p ojec
also highligh ed he alue o in eg a ing hese p edic ions in o a b oade wo k o ce planning amewo k,
ein o cing he s a egic po en ial o da a-d i en HR p ac ices.
The emainde o his chap e is s uc u ed in wo pa s: Sec ion 6.1 p esen s a c i ical
assessmen o he model ou comes, linking hem o s a egic HR decision-making needs, while Sec ion
6.2 p oposes conc e e di ec ions o u u e de elopmen .
6.1 C i ical Analysis S a egic Assessmen and C i ical Re lec ion
Be o e p esen ing a c i ical assessmen o he models de eloped in his disse a ion, i is essen ial
o e isi he b oade con ex , he s a egic mo i a ion o he wo k and he esea ch ques ions ha guided
his in es iga ion. The p ojec se ou o explo e how p edic i e modelling and People Analy ics could be
applied o add ess a pe sis en challenge in he e ail sec o : high olun a y u no e . The objec i e was
wo old — o o ecas employee exi s o e one-yea and i e-yea ho izons and o use hese insigh s as
inpu s o wo k o ce planning a bo h ac ical and s a egic le els. In doing so, he p ojec con ibu ed no
only by de eloping p edic i e models, bu also by demons a ing how such ools can guide HR decision-
making aligned wi h di e en planning ho izons.
This sec ion builds on hose con ibu ions by examining he models’ pe o mance, iden i ying key
limi a ions and assessing hei b oade u ili y o wo k o ce planning. The analysis conside s bo h
me hodological aspec s — such as a iable s uc u e and model calib a ion — and s a egic implica ions
o o ganiza ional decision-making. This c i ical analysis e lec s bo h on he achie emen o he ini ial
objec i es and on he answe s p o ided o RQ1 – “Wha a e he main indi idual and con ex ual ac o s
ha in luence olun a y employee u no e in he e ail sec o ?” – and RQ2 – “How can p edic i e
analy ics be used o o ecas olun a y u no e and suppo s a egic wo k o ce planning in high- u no e
en i onmen s?”.
82
The esul s o bo h p edic i e models p o ide ele an insigh s in o hei applicabili y, while also
e ealing limi a ions ha wa an ca e ul conside a ion. The one-yea model demons a es high
disc imina o y powe and s ong calib a ion, al hough i s absolu e e o a he indi idual le el emains
non-negligible. This limi a ion, howe e , is subs an ially mi iga ed when applied o agg ega ed o ecas s
— especially when segmen ed by Dis ic , S o e Type and Role Clus e — aligning be e wi h he g anula i y
equi ed o wo k o ce planning decisions. The ela i ely modes pe o mance gains obse ed om
in oducing
pe sonas
aise ques ions abou he beha io al cohe ence o hese g oupings. I no e lec i e
o dis inc u no e pa e ns, such segmen a ion may in oduce noise a he han imp o e p edic i e
accu acy.
The i e-yea model, while inhe en ly less p ecise due o i s ex ended o ecas ho izon and he
absence o indi idual-le el da a, s ill deli e s calib a ed p edic ions a he g oup le el. Howe e , he sligh
pe o mance gain om
pe sonas
— along wi h an in e sion in p edic ion bias — sugges s ha he cu en
pe sona
s uc u e may no be well sui ed o long- e m o ecas ing.
Alongside hese insigh s, wo s uc u al limi a ions may be con ibu ing o educed p edic i e
pe o mance. Fi s , he s ong collinea i y be ween age and senio i y could hinde he model’s abili y o
isola e ea ly-ca ee u no e pa e ns. Second, a iables such as Pe o mance and Po en ial exhibi e y
low p edic i e impo ance, likely due o hei skewed dis ibu ions. Al hough included in he
pe sona
amewo k due o hei s a egic ele ance, hese a ibu es added limi ed disc imina i e powe and may
ha e weakened he segmen a ion logic. These aspec s we e no explici ly es ed in his s udy bu ep esen
a enues o u u e e inemen .
Despi e hese challenges, bo h models achie ed good p obabili y calib a ion, con i ming hei
eliabili y o use in scena io analysis and wo k o ce op imiza ion. While he Mean Absolu e E o emains
high a he indi idual le el, p edic ion e o s end o cancel ou in agg ega e — as e lec ed by consis en ly
lowe BIAS — making he models well sui ed o s a egic planning a highe le els o agg ega ion.
F om a s a egic s andpoin , he models’ dis inc oles e lec hei complemen a y alue. The
one-yea model suppo s sho - e m, ac ionable planning, g ounded in in e p e able a iables like
compensa ion and enu e. In con as , he i e-yea model cap u es b oade ends by inco po a ing
ex e nal indica o s such as minimum wage and unemploymen , suppo ing long- e m scena io simula ion.
Toge he , hey enable a mul i-ho izon planning app oach ha aligns immedia e ac ions wi h longe - e m
s a egic o esigh .
Al hough no ye con i med as ac i ely deployed in a eal- ime decision-making en i onmen , he
models ha e al eady been in eg a ed in o a wo k o ce op imiza ion model wi hin he o ganiza ion. This
83
ensu es hey a e eadily a ailable o p ac ical use and ha e he capaci y o suppo wo k o ce planning
h ough scena io-based assessmen s.
Building on hese insigh s, i is now possible o di ec ly add ess he esea ch ques ions ha
guided his disse a ion. In esponse o RQ1, he analysis con i med he cen al ole o compensa ion,
senio i y and b oade mac oeconomic ac o s in in luencing olun a y u no e . Fo RQ2, he de elopmen
and e alua ion o wo machine lea ning models demons a ed ha p edic i e analy ics can e ec i ely
an icipa e u no e ends and in o m wo k o ce planning s a egies ac oss di e en ime ames.
Beyond his analy ical pe spec i e, h ee key con ibu ions eme ge om he wo k. Fi s , i
in oduces a p edic i e amewo k ailo ed o he e ail sec o ’s speci ic challenges. Second, i
demons a es how p edic i e ou pu s can be inco po a ed in o wo k o ce planning ools o suppo bo h
ope a ional and s a egic decisions. Thi d, i p o ides a e lec i e assessmen o me hodological and
s uc u al limi a ions, o e ing clea di ec ions o imp o emen and u u e de elopmen .
Ul ima ely, his wo k ad ances he shi om eac i e o p oac i e HR managemen . By
embedding p edic i e analy ics in o wo k o ce planning, o ganiza ions can be e an icipa e u u e
wo k o ce needs and align hi ing, mobili y and e en ion s a egies acco dingly. This posi ions HR eams
o make mo e in o med decisions, while s eng hening o ganiza ional esilience in dynamic en i onmen s
like e ail.
The nex sec ion ou lines a se o sho - and long- e m ecommenda ions o enhance he models
and ex end hei impac .
6.2 Fu u e Resea ch and De elopmen
Gi en he dynamic na u e o wo k o ce beha iou s and o ganiza ional needs, he con inuous
e inemen o p edic i e models is essen ial. Ra he han p oposing a single lis o u u e imp o emen s,
his sec ion o ganizes ecommenda ions in o wo dis inc ime ho izons — sho - e m and medium/long-
e m — o be e e lec hei complexi y, easibili y and expec ed impac .
Sho - e m de elopmen s ocus on op imizing he cu en amewo k, enhancing inpu s uc u es
and imp o ing model in e p e abili y and alignmen wi h decision-making needs. These a e ac ions ha
can be pu sued using a ailable da a and ools, wi h ela i ely low implemen a ion isk.
In con as , medium o long- e m de elopmen s aim o ex end he capabili ies o he model by
in eg a ing ad anced o ecas ing me hods, explo ing al e na i e modeling echniques and enabling iche
scena io simula ions. These di ec ions in ol e a highe le el o unce ain y o equi e s uc u al changes
in da a collec ion bu o e he po en ial o signi ican s a egic alue o e ime.
84
6.2.1 Sho -Te m Imp o emen s
Pe sonas
Re ision
A key oppo uni y o imp o emen lies in how
pe sonas
a e de ined. In he cu en model,
pe sonas
a e p e-s uc u ed based on a se o manually selec ed a iables such as con ac ype, wo kload
ype, academic quali ica ions, pe o mance, po en ial, age and enu e. While hese a iables we e chosen
o hei s a egic ele ance and p ac ical a ailabili y, he way in which hey a e segmen ed — pa icula ly
he nume ical ones — may no e ec i ely cap u e eal beha io al di e ences in u no e .
Fo ins ance, a bi a y disc e iza ion o con inuous a iables like age o enu e (e.g., c ea ing bins
such as “unde 25,” “25–35,” “35–45”) may g oup oge he employees who do no sha e simila
pa e ns o u no e , while sepa a ing hose who do. This may esul in poo ly de ined
pe sonas
ha dilu e
he p edic i e powe o he model.
To add ess his, i is ecommended o explo e unsupe ised clus e ing echniques (e.g., K-means,
DBSCAN, hie a chical clus e ing), which can conside bo h ca ego ical and nume ical a iables in hei
aw o m. These app oaches can au oma ically iden i y na u ally occu ing pa e ns in he da a — including
op imal g oupings o con inuous a iables — esul ing in
pe sonas
ha a e mo e cohesi e in e ms o
ac ual u no e beha io . By allowing he da a o guide bo h he selec ion o a iables and he way hey
a e g ouped, he esul ing
pe sonas
may be e e lec he ue he e ogenei y o he wo k o ce, inc easing
he model’s abili y o gene alize and enhancing he e ec i eness o he op imiza ion amewo k.
Imp o emen o Pe o mance and Po en ial E alua ions
The cu en use o
Pe o mance
and
Po en ial
a iables is signi ican ly limi ed by he way hese
a ibu es a e dis ibu ed in he da ase . As shown in he igu es, bo h a iables a e hea ily skewed:
• Pe o mance: Mo e han 64% o employees ecei e a sco e o 3, ollowed by 26% a le el 4. The
emaining le els (1, 2 and 5) a e hea ily unde ep esen ed (Figu e 12).
• Po en ial: App oxima ely 80% o employees a e a ed as le el 1, wi h only 16% in le el 2 and less
han 5% in le el 3 (Figu e 13).
This ex eme concen a ion a ound one o wo alues g ea ly educes he disc imina i e powe o
hese a iables. Fo example, i nea ly all employees a e assigned a po en ial sco e o 1, he model canno
e ec i ely di e en ia e u no e pa e ns ac oss po en ial le els — limi ing i s abili y o ex ac meaning ul
insigh s om his ea u e.
In he sho e m, no maliza ion o escaling echniques such as z-sco e ans o ma ion o
quan ile scaling a e no sui able in his case, as he da a a e o dinal, ca ego ical and highly imbalanced.
85
Ins ead, al e na i e p ep ocessing me hods could be explo ed, such as smoo hing o weigh ing
mechanisms o sligh ly inc ease he in luence o unde ep esen ed classes wi hou a i icially dis o ing
he da ase ’s s uc u e.
Howe e , he mos impac ul imp o emen lies in ede ining how hese a iables a e e alua ed.
Collabo a ing wi h HR o e ise he cu en e alua ion amewo k is ecommended, wi h he goal o
p oducing mo e in o ma i e and balanced dis ibu ions. Po en ial s eps include:
• In oducing mo e g anula a ing sys ems (e.g., 1 o 10 scale);
• S anda dizing calib a ion c i e ia o ensu e consis en e alua ions ac oss eams and manage s;
• P o iding clea guidelines and benchma ks o educe a ing in la ion o clus e ing a ound he
median;
• Combining subjec i e assessmen s wi h objec i e me ics such as goal achie emen , skills
de elopmen o mobili y his o y.
By imp o ing bo h he scale and he consis ency o hese e alua ions, hese ea u es could
become signi ican ly mo e aluable o u no e p edic ion. In addi ion o enhancing model pe o mance,
such imp o emen s could os e g ea e ai ness and anspa ency in employee assessmen p ac ices —
b inging added alue o bo h he p edic i e sys em and HR decision-making as a whole.
Analysis o Collinea i y Be ween Age and Senio i y
Du ing he explo a o y analysis o he da ase , a s ong co ela ion was obse ed be ween he
a iables
age
and
senio i y
. This ela ionship aised ea ly conce ns abou mul icollinea i y and was la e
suppo ed by he model’s pe o mance ac oss di e en employee segmen s: while he model showed
high accu acy in p edic ing u no e by
age
, i s uggled o do so by
senio i y
— pa icula ly
unde es ima ing u no e in ea ly-ca ee g oups (e.g., 1–3 yea s and 3–6 yea s o enu e). Gi en ha
hese wo a iables a e s uc u ally ela ed, such inconsis ency sugges s ha he model may be ailing o
disen angle hei indi idual e ec s.
This beha io may indica e ha
age
is abso bing mos o he explana o y powe o
senio i y
,
leading o a edundancy ha comp omises in e p e abili y and po en ially limi s p edic i e pe o mance
in segmen s whe e age and enu e do no align pe ec ly. To add ess his issue, a o mal mul icollinea i y
analysis should be conduc ed — s a ing wi h co ela ion ma ices and a iance in la ion ac o (VIF)
calcula ions.
I high mul icollinea i y is con i med, se e al s a egies can be conside ed:
• Excluding one o he a iables o simpli y he model;
86
• Applying dimensionali y educ ion echniques (e.g., p incipal componen analysis) o cap u e
sha ed a iance;
• C ea ing enginee ed ea u es ha exp ess hei in e ac ion mo e e ec i ely (e.g., age- o- enu e
a io o enu e adjus ed o age g oup).
This analysis should no be es ic ed o age and senio i y. Simila checks should be ex ended o
o he po en ially co ela ed ea u e pai s. The insigh s om hese analyses can guide u he model
e inemen , helping o imp o e bo h s abili y and in e p e abili y by ensu ing ha each a iable adds
unique, non- edundan alue o he model’s p edic ions.
S eng hening Local In e p e abili y wi h SHAP
While he global ea u e impo ance analysis p o ides aluable insigh s in o he gene al beha iou
o he model, local in e p e abili y — unde s anding why he model made a speci ic p edic ion o a speci ic
g oup — is equally impo an , especially o ope a ional decision-making a he egional le el.
To achie e his, i is ecommended o implemen SHAP (SHapley Addi i e exPlana ions) a he g oup le el
— pa icula ly o combina ions such as dis ic x s o e ype x ole clus e x
pe sona
. SHAP alues can
help explain how each ea u e con ibu es (posi i ely o nega i ely) o he p edic ed u no e o a speci ic
g oup, o e ing a mo e nuanced and anspa en iew han global ea u e impo ance alone.
Fo ins ance, in one egion, high u no e migh be d i en by low wages, while in ano he i migh
be mo e a ec ed by he p e alence o ixed- e m con ac s. By highligh ing hese local dynamics, SHAP
explana ions empowe egional HR manage s o design a ge ed e en ion s a egies and us he model’s
ou pu s, ul ima ely inc easing he impac and adop ion o da a-d i en decisions.
Fu u e Adjus men o Sys ema ic Bias
Despi e he s ong a e age pe o mance, bo h he one-yea and i e-yea models display
sys ema ic bias in ce ain segmen s — consis en ly o e es ima ing o unde es ima ing u no e . These
pa e ns can be add essed by implemen ing a ligh weigh bias co ec ion mechanism based on his o ical
e o s.
Such a mechanism could in ol e calcula ing he a e age e o o each segmen o e ime (e.g.,
by dis ic x s o e ype x ole clus e x
pe sona
) and hen applying a co ec ion ac o o u u e p edic ions
o hose g oups. This should be done cau iously, as o e ly agg essi e adjus men s could lead o
o e i ing o ins abili y. Howe e , i applied conse a i ely and moni o ed egula ly, i could imp o e
alignmen be ween p edic ed and ac ual u no e , especially o ecu ing edge cases.
87
Addi ionally, his eedback loop could be embedded in o he model’s li ecycle — igge ing mino
upda es o ale s when bias exceeds p ede ined h esholds.
Mo e F equen Model Upda es
Cu en ly, he model is upda ed on an annual basis, ollowing he yea ly da a e esh. Howe e ,
labou dynamics and ex e nal ac o s can shi mo e apidly, especially du ing pe iods o o ganiza ional
change, economic ola ili y o policy e o m.
I is he e o e sugges ed o e alua e he easibili y o semi-annual upda es, pa icula ly o he
one-yea model, which is mo e sensi i e o ecen ends. These upda es would no necessa ily equi e
e aining he en i e model — in some cases, inc emen al lea ning app oaches o as ecalib a ion o
ce ain componen s could su ice.
By inc easing he e esh a e, he models could espond mo e quickly o changing condi ions
and a oid he lag ha comes wi h ou da ed aining da a. This is pa icula ly ele an in as -mo ing
en i onmen s such as e ail, whe e u no e pa e ns can shi signi ican ly in a ma e o mon hs.
6.2.2 Medium/Long-Te m Imp o emen s
In eg a ion o In e nal Mobili y Dynamics
A p esen , he u no e p edic ion model ope a es on an annual basis, meaning ha i only
conside s indi iduals o
pe sonas
who emain in he company un il he end o he yea o p edic hei
u no e o he ollowing yea . Consequen ly, employees who en e and lea e wi hin he same yea a e
excluded om he model, as hey do no mee he c i e ia o he u no e s a us a he yea -end. This
limi a ion means ha any in a-yea dynamics — such as seasonal luc ua ions in u no e , especially
common in e ail en i onmen s — a e no being cap u ed.
The absence o seasonal u no e , which can be signi ican ly highe in ce ain pe iods (such as
summe o holiday seasons), ep esen s a gap in he model’s p edic i e capabili y. While he cu en
model ocuses solely on p edic ing u no e based on he s a us a yea -end, inco po a ing u no e du ing
he yea , including en ies and exi s wi hin he same pe iod, could o e a mo e comp ehensi e iew o
wo k o ce dynamics. Such an imp o emen would make he model mo e e lec i e o eal-wo ld employee
mo emen , pa icula ly in indus ies like e ail whe e seasonal u no e is a common end.
88
Expansion o Ex e nal Da a Sou ces
While he i e-yea model al eady includes mac oeconomic a iables such as minimum wage and
unemploymen a e, i s p edic i e po en ial could be enhanced by inco po a ing addi ional ex e nal da a
sou ces. These may include egion-speci ic labou demand indica o s, in la ion-adjus ed cos o li ing
indexes, eal es a e p ices o e en commu ing and anspo a ion accessibili y me ics.
By b oadening he con ex ual scope, he model can be e cap u e a ia ions in u no e d i e s
ac oss egions and ime pe iods. Fo example, a dis ic wi h limi ed public anspo a ion and ising li ing
cos s migh exhibi highe u no e e en i in e nal compensa ion emains s able. In eg a ing hese
ex e nal da ase s would also suppo mo e nuanced scena io analyses and enhance he ealism o long-
e m wo k o ce planning.
E hical and Fai ness Conside a ions
As p edic i e models inc easingly suppo s a egic HR decisions, i becomes impo an o ensu e
hey do no unin en ionally ein o ce inequi ies. Al hough he cu en models do no include sensi i e
a ibu es such as gende o e hnici y, po en ial p oxy bias may s ill exis h ough co ela ed a iables like
age o con ac ype. Fu u e de elopmen s should include ai ness e alua ions o p omo e e hical and
equi able use o p edic i e analy ics in wo k o ce planning.
Benchma king wi h Al e na i e A chi ec u es
While XGBoos has p o en o be a obus and in e p e able algo i hm, he con inuous e olu ion
o p edic i e modelling in HR analy ics opens he doo o explo ing al e na i e app oaches. In he medium
o long e m, i may be wo hwhile o benchma k he cu en model agains bo h mo e ad anced
a chi ec u es and simple ones, in o de o assess whe he di e en modelling s a egies o e g ea e
s abili y o adap abili y in unce ain long- e m o ecas ing scena ios. The p io i y should emain on
balancing p edic i e pe o mance wi h in e p e abili y and obus ness.
Inco po a ing Exi In e iew and Quali a i e Da a
One o he limi a ions o he cu en model is i s eliance on s uc u ed, quan i a i e a iables.
Howe e , quali a i e signals — such as exi in e iew no es, employee engagemen su eys o w i en
pe o mance e iews — may con ain aluable insigh s abou he unde lying easons o u no e .
Na u al Language P ocessing (NLP) echniques could be used o ex ac sen imen , hema ic
pa e ns o beha iou al cues om hese uns uc u ed sou ces. These insigh s could hen be inco po a ed
89
as addi ional ea u es o used o alida e and con ex ualize model p edic ions. While he in eg a ion o
such da a equi es ca e ul e hical and p i acy conside a ions, i o e s a powe ul a enue o complemen
nume ical models wi h human-cen ic unde s anding.
Ope a ional Feedback Loops and Con inuous Lea ning
A long- e m goal o he modelling sys em is o embed a eal- ime eedback mechanism in o he
HR decision-making p ocess. Cu en ly, p edic ions a e made and in e p e ed ex e nally om he
ou comes hey seek o in luence. By closing his loop — o ins ance, by eco ding how well he model's
p edic ions aligned wi h eali y a e wo k o ce planning ac ions we e implemen ed — he model could
lea n con inuously om i s en i onmen .
This would allow o adap i e ecalib a ion, as e de ec ion o concep d i and a mo e p oac i e
app oach o wo k o ce planning. While such sys ems equi e in es men in in as uc u e and go e nance,
hey ep esen he nex s ep in making p edic i e analy ics no jus a o ecas ing ool, bu a co e
componen o s a egic HR ope a ions.
96
APPENDIX
Appendix A – Suppo ing Da a and Ex ended Analysis
Figu e 36 demons a es he ela ionship be ween ma i al s a us and u no e a e, e ealing
no iceable di e ences be ween single and ma ied employees. While single employees make up he
majo i y o he wo k o ce a 66.49%, ma ied employees accoun o 33.51%. Despi e his, u no e is
signi ican ly highe among single employees, a 13.22%, compa ed o jus 4.51% o ma ied employees.
This pa e n sugges s ha single employees a e mo e inclined o lea e he company han hei
ma ied colleagues. Se e al ac o s may con ibu e o his dispa i y. Single employees o en ha e ewe
obliga ions ying hem o a speci ic job, allowing o g ea e mobili y and lexibili y o seek new
oppo uni ies. Addi ionally, hey may be a an ea lie s age in hei ca ee s, whe e job ansi ions a e mo e
equen as hey explo e ca ee ad ancemen s, sala y imp o emen s o di e en p o essional
expe iences.
In con as , ma ied employees end o show g ea e job s abili y, likely in luenced by inancial
esponsibili ies, amily commi men s and a s onge p e e ence o long- e m secu i y. Thei p o essional
choices may be guided by he need o s abili y, educing he likelihood o olun a y u no e .
Figu e 36 - Wo k o ce Dis ibu ion and Tu no e Ra e by Ma i al S a us
Figu e 37 highligh s he dis ibu ion o employees ac oss di e en s o e ypes based on hei
senio i y. This isualiza ion complemen s Figu e 4 by p o iding addi ional insigh s in o how wo k o ce
composi ion a ies ac oss s o e o ma s. The key akeaway is ha con enience s o es a e o e whelmingly
popula ed by employees wi h lowe senio i y, while hype ma ke s concen a e he majo i y o employees
97
wi h o e 10 yea s o expe ience. Supe ma ke s, on he o he hand, p esen a mo e balanced mix o all
senio i y le els.
No ably, he e a e almos no employees wi h mo e han 10 yea s o senio i y in con enience
s o es, ein o cing he idea ha hese loca ions a e p ima ily s a ed by newe employees. In con as ,
hype ma ke s ha e he highes p opo ion o long- enu ed employees, sugges ing ha hey o e g ea e
long- e m ca ee s abili y o p og ession oppo uni ies. Supe ma ke s se e as an in e media y ca ego y,
accommoda ing bo h newe and mo e expe ienced employees wi hou a s ong dominance o any
pa icula senio i y g oup.
This dis ibu ion may indica e ha con enience s o es se e as an en y poin o employees who
la e ansi ion o o he s o e ypes, while hype ma ke s e ain a mo e expe ienced wo k o ce, possibly
due o be e ca ee g ow h oppo uni ies, enhanced job secu i y o di e ences in s o e ope a ions ha
in luence long- e m e en ion.
Figu e 37 - S o e Type Dis ibu ion by Senio i y
Figu e 38 highligh s a clea upwa d end in a e age employee age as senio i y inc eases.
Employees in he ea ly yea s o enu e end o be younge , while hose wi h longe enu e a e signi ican ly
olde on a e age. The p og ession is g adual, showing a s eady inc ease in age as employees gain mo e
yea s in he company.
This pa e n sugges s ha wo k o ce aging is na u ally linked o enu e, wi h younge employees
ei he ansi ioning o highe senio i y le els o e ime o lea ing he company ea ly. These indings
ein o ce he ela ionship obse ed in Figu e 6 and Figu e 7, whe e senio i y and age we e sepa a ely
analyzed. By di ec ly linking he wo a iables, Figu e 38 con i ms ha olde employees end o hold
highe - enu e posi ions, p o iding u he alida ion o he ends p e iously discussed.
98
Figu e 38 - Age Dis ibu ion by Senio i y
Figu e 39 illus a es he ela ionship be ween age and con ac ype, showing a clea end whe e
ixed- e m con ac s a e p edominan ly assigned o younge employees, while pe manen con ac s
become he no m as employees g ow olde . Mos employees unde 25 yea s old a e on ixed- e m
con ac s, wi h his being especially e iden in he ≤ 20 yea s g oup, whe e o e 95% o employees hold
empo a y posi ions. As age inc eases, he sha e o pe manen con ac s ises p og essi ely, su passing
ixed- e m con ac s a ound he 25-35 yea s ca ego y. By he 35-50 yea s age g oup, pe manen con ac s
become he o e whelming majo i y and among employees o e 50, almos all ha e pe manen
employmen .
This dis ibu ion aligns wi h he u no e pa e ns obse ed in Figu e 8, whe e ixed- e m employees
exhibi a signi ican ly highe olun a y u no e a e. Gi en ha younge employees a e mo e likely o hold
ixed- e m con ac s, he highe u no e among his g oup may be mo e e lec i e o age and ca ee
s age mobili y a he han jus con ac ype alone. Employees in ea ly ca ee phases may be mo e likely
o seek new oppo uni ies, change jobs o ansi ion in o ull- ime oles, con ibu ing o he highe u no e
a e obse ed in ixed- e m con ac s.
Thus, a he han ixed- e m con ac s being he di ec cause o highe u no e , his g aph sugges s
ha he endency o younge employees o ha e empo a y con ac s — and hei highe mobili y in he
job ma ke —may be key ac o s in explaining he obse ed u no e di e ences.
99
Figu e 39 - Con ac Type Dis ibu ion by Age
Figu e 40 es ablishes he connec ion be ween wo kload ype and con ac ype, ein o cing he
pa e ns obse ed in Figu e 8 and Figu e 11, which highligh ed he impac o bo h con ac s abili y and
wo kload ype on olun a y u no e . The cha e eals ha ull- ime employees o e whelmingly hold
pe manen con ac s (82.44%), whe eas in pa - ime oles, he e is a mo e e en dis ibu ion, wi h 58.13%
on ixed- e m con ac s and 41.87% on pe manen con ac s.
Gi en ha ixed- e m and pa - ime employees bo h exhibi highe u no e a es, as seen in
Figu e 8 and Figu e 11, his isualiza ion p o ides addi ional e idence ha wo kload ype and con ac
s abili y a e closely linked and may be ac ing oge he as key ac o s in luencing olun a y u no e . Since
pa - ime employees a e mo e likely han ull- ime employees o ha e ixed- e m con ac s, hei highe
u no e may no be due solely o hei wo king hou s bu also o he ac ha hey hold less s able
con ac ual ag eemen s, making olun a y exi s mo e equen .
This insigh helps explain why employees wi h pa - ime and ixed- e m con ac s lea e mo e o en—
no only due o hei wo kload ype o con ac ype indi idually, bu because hese wo ac o s o en
o e lap. This ein o ces a pa e n o highe mobili y and job ansi ions among less s able employmen
ca ego ies.
100
Figu e 40 - Con ac Type Dis ibu ion by Wo kload Type
Figu e 41 p esen s he dis ibu ion o employees ac oss di e en ole clus e s and hei co esponding
a e age s anda dized wage. A clea dis inc ion eme ges be ween oles wi h a highe p opo ion o
employees and hose wi h highe s anda dized wages.
The Cashie Ope a o and Flow Ope a o posi ions accoun o nea ly hal o he wo k o ce, making
hem he mos p e alen oles. Howe e , hese posi ions, along wi h mos o he on line ope a ional
unc ions, main ain a s anda dized wage le el o 15, indica ing limi ed sala y p og ession despi e hei
p e alence.
In con as , oles ha equi e specialized skills o manage ial esponsibili ies — such as
Main
Supe iso
,
Sec ion Coo dina o
and hose classi ied unde he
O he
ca ego y — exhibi signi ican ly
highe wages, wi h s anda dized alues eaching 29-32. The
O he
ca ego y is pa icula ly no ewo hy,
displaying one o he highes a e age wages despi e ep esen ing a small ac ion o he wo k o ce. This
sugges s ha i comp ises highly specialized o senio posi ions ha , while less common, command
signi ican ly highe compensa ion due o hei s a egic impo ance o he expe ise equi ed.
This deepe b eakdown o sala y dis ibu ion ac oss oles builds upon he insigh s om Figu e 9,
whe e he o e all ela ionship be ween sala y le els and u no e a es was in oduced. A mo e de ailed
examina ion o he
O he
ca ego y and i s wage dis ibu ion is p o ided in in Figu e 42, which ocuses on
speci ic oles wi hin his classi ica ion.
101
Figu e 41 - Wo k o ce Dis ibu ion and S anda dized Wage by Role Clus e
Figu e 42 highligh s a di e se mix o ole clus e s, encompassing bo h high- esponsibili y leade ship
posi ions and mo e specialized o ansi ional oles wi h ewe employees. The key obse a ion is he clea
dispa i y in s anda dized wages, wi h manage ial posi ions s anding ou as he highes -paid oles, while
mos ope a ional o suppo unc ions emain a lowe wage le els.
A pa icula ly no able case is he
S o e Di ec o – Hmk
(Hype ma ke ), which exhibi s a s anda dized
wage o 111, signi ican ly su passing all o he oles. This ein o ces he hie a chical wage s uc u e, whe e
leade ship oles o e seeing la ge and mo e complex s o e o ma s command he highes sala ies.
Simila ly, he
S o e Di ec o – Smk
(Supe ma ke ) and
A ea Coo dina o - Hmk
(Hype ma ke ) also
demons a e ele a ed wages ela i e o hei wo k o ce p opo ion, con i ming he p emium placed on
senio manage ial posi ions.
Addi ionally, i is impo an o highligh ha he highes - anking posi ion wi hin con enience s o es, he
C n (Con enience)
S o e Manage
, has a signi ican ly lowe s anda dized wage (38) compa ed o i s
coun e pa s in supe ma ke s and hype ma ke s. This sugges s ha s o e o ma plays a key ole in
de ining sala y s uc u es, wi h la ge s o es demanding g ea e manage ial complexi y, b oade
esponsibili ies and highe compensa ion le els. The s a k con as be ween con enience s o e leade ship
and supe ma ke /hype ma ke di ec o s u he ein o ces he p og essi e wage di e en ia ion based on
s o e size and ope a ional scale.
In con as , oles such as
Ope a o 360
and
Sales Ope a o
main ain bo h low ep esen a ion and low
s anda dized wages, indica ing ha hese posi ions a e ei he en y-le el, empo a y o equi e ewe
specialized skills. Meanwhile, echnical o ope a ional oles such as
Quali y Con ol Ope a o
,
Cus ome
Suppo Se ice
and
Main enance Technician
occupy a middle g ound, wi h wages sligh ly abo e he base
le el bu s ill well below manage ial oles.
A pa icula ly in e es ing obse a ion is he p esence o
A ea Coo dina o
oles ac oss di e en s o e
ypes (Hype ma ke , Supe ma ke and Con enience). These posi ions exhibi a ying wage le els, wi h
hype ma ke coo dina o s ea ning signi ican ly mo e han hei coun e pa s in smalle s o e o ma s. This
sugges s ha s o e size and complexi y in luence compensa ion, e en wi hin he same hie a chical le el.
102
O e all, his g aph ein o ces he dual na u e o ole clus e s wi hin he company: senio leade ship
oles a e sca ce bu highly compensa ed, while ope a ional, suppo and ansi ional posi ions end o
ha e lowe wages and ep esen a ion. The signi ican wage gap be ween s o e di ec o s and lowe -le el
oles e lec s he company's s uc u ed ca ee p og ession, whe e g ea e manage ial esponsibili y
ansla es in o highe sala ies.
Figu e 42 - Wo k o ce Dis ibu ion and S anda dized Wage by Role Clus e (O he s)
