Evaluating credit default risk in P2P lending: a market maturity perspective

i EVALUATING CREDIT DEFAULT RISK IN P2P LENDING Ri a Se as Celo ico Da Sil a Fialho A Ma ke Ma u i y Pe spec i e Disse a ion p esen ed as pa ial equi emen o ob aining he Mas e ’s deg ee in S a is ics and In o ma ion Managemen ii NOVA In o ma ion Managemen School Ins i u o Supe io de Es a ís ica e Ges ão de In o mação Uni e sidade No a de Lisboa EVALUATING CREDIT DEFAULT RISK IN P2P LENDING by Ri a Se as Celo ico Da Sil a Fialho Disse a ion p esen ed as pa ial equi emen o ob aining he Mas e ’s deg ee in S a is ics and In o ma ion Managemen Ad iso : P o . D . Jo ge Miguel Ven u a B a o, PhD No embe 2021 iii ABSTRACT The Pee - o-Pee (P2P) lending indus y has g own signi ican ly in ecen yea s, owing i s ac ion o a g owing in e es in he ma ke om bo h indi idual bo owe s and lende s alike. Such pla o ms ha e g ea ly bene i ed om he ad en o digi al ans o ma ion and expanding in e ne oo p in ac oss people o all ages and backg ounds. These pla o ms le e age no el ways o in e ac ing o acili a e an al e na i e means o inancing which p o ides mo e oppo uni ies o bo owe s who may, o he wise no ha e access o deb h ough con en ional mechanisms, and mo e in es men oppo uni ies o lende s who may be seeking o di e si y hei po olios. This s udy le e ages a da ase made a ailable by he Lending Club P2P lending pla o m, con aining mo e han 2 million loan en ies, amassed o e mo e han 10 yea s. Being a pionee in he indus y, Lending Club’s ack eco d and da a p o ide a good window in o he inne wo kings o such pla o ms and hei abili y o assess he c edi wo hiness o bo owe s and loan applica ions. We se ou o build upon p io wo k done in his space by unde s anding and analyzing his da ase o iden i y he de e minan s o de aul o loans issued h ough P2P lending pla o ms. Ou analysis is also employed o c ea e a p edic i e model which is hen es ed agains ou da ase . This app oach builds upon p e ious s udies by ou lining an end- o-end p ocess o analyze and assess a pla o m’s abili y o adequa ely p edic c edi de aul isk. We ha e ound ha , in alignmen wi h p io wo k, such pla o ms a e indeed able o adequa ely assess c edi de aul isk, in he way ha g ades a e assigned o indi idual loans. The logis ic eg ession model which we ha e buil has also yielded good esul s in p edic ing de aul ed loans, while exhibi ing medioc e pe o mance in classi ying ully epaid loans as likely cases o de aul . KEYWORDS C edi Risk; C edi De aul ; De aul Risk; P2P Lending; Lending Club; Logis ic Reg ession; Da a Analysis; De e minan s o De aul i INDEX 1 In oduc ion ................................................................................................................ 8 1.1 Lending Club ......................................................................................................... 9 2 Li e a u e e iew ....................................................................................................... 13 2.1 Online P2P lending his o y and cha ac e iza ion ............................................... 13 2.2 Risks o online P2P lending ma ke s – he good and he bad ............................ 13 2.3 De e minan s o de aul in online P2P lending ma ke s .................................... 15 2.4 Risk sco ing in online P2P lending ma ke s ........................................................ 15 3 Me hodology ............................................................................................................. 17 3.1 The Model - Logis ic Reg ession ......................................................................... 17 3.2 Pe o mance Measu emen s o he Model ....................................................... 18 4 Da a ........................................................................................................................... 20 4.1 Da a Desc ip ion ................................................................................................. 20 4.2 Da a Re inemen ................................................................................................ 21 4.3 Explo a o y Da a Analysis .................................................................................. 24 4.4 Final Va iables .................................................................................................... 28 5 Resul s and discussion ............................................................................................... 30 5.1 Model and Resul s .............................................................................................. 30 5.2 Model Pe o mance ........................................................................................... 33 6 Conclusions ............................................................................................................... 35 6.1 Limi a ions and Fu u e Wo k ............................................................................. 36 7 Bibliog aphy .............................................................................................................. 37 8 Appendix ................................................................................................................... 42 8.1 Appendix A ......................................................................................................... 42 8.2 Appendix B ......................................................................................................... 47 8.3 Appendix C ......................................................................................................... 48 LIST OF FIGURES Figu e 1 – Lending Club’s Business Model ................................................................................ 9 Figu e 2 - Numbe o loans issued by yea .............................................................................. 20 Figu e 3 - Numbe o obse a ions by employmen leng h .................................................... 22 Figu e 4 - Dis ibu ion o annual income ................................................................................ 25 Figu e 5 - D i ou lie s .............................................................................................................. 25 Figu e 6 - Rela ionship be ween in e es a e and subg ade .................................................. 26 Figu e 7 - Co ela ion Ma ix o nume ical a iables .............................................................. 27 Figu e 8 - ROC Cu e ............................................................................................................... 34 i LIST OF TABLES Table 1 - Con usion Ma ix ...................................................................................................... 18 Table 2 - Summa y o he employmen leng h a iable .......................................................... 22 Table 3 - Numbe o loans by loan s a us ............................................................................... 23 Table 4 - Summa y o he annual income a iable ................................................................. 24 Table 5 - Dis ibu ion o loans by home owne ship ................................................................ 27 Table 6 - Logis ic Reg ession esul s ....................................................................................... 31 Table 7 - Con usion Ma ix ...................................................................................................... 33 Table 8 - Model Pe o mance Measu es ................................................................................ 33 Table 9 - All Lending Club a iables wi h desc ip ion ............................................................. 42 Table 10 - Co ela ion ma ix o LC a iables .......................................................................... 47 Table 11 - Desc ip i e S a is ics o Nume ical Va iables ......................................................... 48 Table 12 - Desc ip i e S a is ics o Ca ego ical Va iables ....................................................... 49 ii LIST OF ABBREVIATIONS AND ACRONYMS ANNs A i icial Neu al Ne wo ks AUC A ea Unde he ROC Cu e BME Bayesian Model Ensembles DA Disc imina e Analysis DT Decision T ees DTI Deb o Income FPR False Posi i e Ra e GA Gene ic and E olu iona y Algo i hms LR Logis ic Reg ession MLM Maximum Likelihood Me hod NN Neu al Ne wo ks P2P Pee - o-pee OR Odds Ra io RF Random Fo es ROC Recei e Ope a ing Cha ac e is ic cu e SEC Secu i ies and Exchange Commission SVM Suppo Vec o Machine TPR T ue Posi i e Ra e 8 1 INTRODUCTION In mode n his o y, lending is conside ed o be one o he key ac i i ies ca ied ou by inancial ins i u ions. I is h ough he ins umen o c edi ha a ious economic ac i i ies see hei sp ead and di e si ica ion, om p oduc ion o consump ion. I is h ough c edi ha businesses and indi iduals uel hei needs o g ow and imp o e. In he ac i i y o “lending” money, wo pa ies en e a ansac ion in which a lende “lends” money o a bo owe , wi h a p omise o g adual epaymen . Rega dless o he e ms o epaymen , he e is always a isk associa ed wi h he bo owe no being able o pay back he loan. This isk is classi ied as “c edi isk” and is, hence, a measu e o he likelihood ha a bo owe may de aul on an issued loan, hus ailing o mee he e ms de ined in he con ac en e ed by bo h pa ies. Gi en he na u e and unce ain y o he c edi ma ke and conside ing he isks inhe en o he money- lending ac i i y, inancial ins i u ions ha e long sough o bo h de elop and imp o e he mechanisms h ough which hey e and check he quali y o bo owe s o whom hey lend money, as a means o educing c edi isk and, ul ima ely, making back hei ini ial in es men . The inancial c isis o 2007-2010 had a widesp ead and global impac on he en i e inancial indus y and as ly a ec ed o he sec o s o he economy, leading o a educ ion in he possibili y o ecou se o adi ional c edi by consume s and small businesses and dis us and dissa is ac ion wi h comme cial banks and o he es ablished inancial ins i u ions. The ad en o he in e ne and he inc ease in popula i y o web se ices and new business models based on le e aging he bene i s o in e connec i i y and as , ubiqui ous access o he in e ne , coupled wi h he a o emen ioned dis us and discon en wi h adi ional inancial ins i u ions and in e media ies, u he accele a ed he ise o al e na i e inance in he con ex o a globalized wo ld economy. Al e na i e inancial se ices compose an a ay o inancial se ices o e ed by p o ide s ha may ope a e ou side o he umb ella o egula ed ins i u ions. Despi e many o he p oduc s and se ices p o ided by hem no being “al e na i e”, bu a he he same as, o simila o hose o , adi ional p o ide s, ins i u ions in his segmen a e o en cha ac e ized by ha ing mo e inhe en isk o hei ope a ions and by execu ing adi ional unc ions/se ices ia al e na i e means o in e ac ion (e.g. le e aging he In e ne o connec inancial pee s in a ansac ion) (B adley e al., 2009). Pee - o-pee (P2P) lending is he p ac ice in which indi iduals lend money o o he indi iduals h ough, gene ally, an online pla o m ha di ec ly connec s lende s wi h bo owe s wi hou he in ol emen o a adi ional inancial in e media y. Whe e in con en ional lending a inancial ins i u ion, such as a bank, would lend money o an indi idual, equen ly aking some ype o colla e al o secu e i (such as a house o an au omobile), his new o m o lending complemen s he ma ke by opening lending as an in es men ehicle di ec ly a ailable o indi iduals, and expanding he oppo uni ies p o ided o a wide ange o bo owe s wi h, po en ially less es ic ions o cons ain s (such as age, c edi wo hiness and employabili y). P2P lending pla o ms o e simple egis a ion p ocesses o bo h bo owe s and lende s alike, allowing hem o submi de ails o hei c edi needs and unding in e es , espec i ely. A e 9 conduc ing he necessa y due diligence, hey connec bo owe s and lende s o ill loan eques s and p o ide unding ac oss a ange o isk le els (based on he c edi a ing p o ided o each loan). The i s su acing o p ac ical, gene ally accep ed and unc ional pla o ms aces back o Zopa, a B i ish P2P lending company, ounded in 2005 in he UK. Soon a e i , o he companies ollowed he same model, wi h many non- iable candida es and al e na i es coming o be. O he success ul pla o ms we e also s a ed in he US, wi h P ospe (2006) and Lending Club (2007) being he mos no able men ions. They a e cu en ly he wo la ges P2P companies in he US, espec i ely holding igu es o US$38 billion in issued loans and o e 2.5 million cus ome s o Lending Club and US$13 billion in issued loans and o e 820 housand cus ome s o P ospe (a he ime o w i ing). Despi e he espec able igu es p esen ed abo e, and he g ow h end e idenced in he sec o (which is expec ed o g ow o a US$897.85 billion ma ke by 2024), he P2P lending segmen s ill holds a conside ably small pe cen age o all consume c edi issued, when compa ed o he global igu es o US$3.9 illion. (T anspa ency Ma ke Resea ch, 2016) P2P lending pla o ms do, none heless, me i he a en ion hey ecei e. They con inue he end o disin e media ion ha online business models ha e es ablished in mode n-day In e ne and, hus, p o ide di ec access o unsecu ed consume c edi as an al e na i e in es men ehicle o indi idual lende s and do so a a much educed “in e media ion” cos due o de ia ing om he s anda d p ac ice o adi ional c edi o ganiza ions. They also democ a ize access o consume c edi , by employing he a o emen ioned model, and b ing i o he each o he o he wise non-c edi wo hy, by applying he necessa y sa egua ds and adequa ely p icing isk. (He zens ein, And ews, Dholakia, & Lyand es, 2008) 1.1 LENDING CLUB To allow o be e comp ehension o he da a which will be used in la e chap e s, i is impo an ha we i s de elop an unde s anding o he subjec unde s udy (Lending Club), by e iewing how i ope a es and wha ex e nali ies condi ion i s ope a ion. Figu e 1 – Lending Club’s Business Model Sou ce: Lending Club Co p Fo m 10-K o Fiscal Yea Ended Decembe 31, 2014. Secu i ies and Exchange Commission. 16 To assess c edi isk, in de eloped ma ke s lende s ypically conside his o ical loan applica ion and loan pe o mance da a collec ed egula ly om a small numbe o sou ces based on long-s anding banking and c edi ela ionships o de elop c edi -sco ing models. T adi ional c edi -sco ing models applying single-pe iod classi ica ion echniques (e.g., logi , p obi ) o classi y c edi cus ome s in o di e en isk g oups and o es ima e he p obabili y o de aul a e s ill he mos popula da a mining echniques used in he indus y (Chamboko & B a o, 2016, 2019a,b, 2020). Indi idual classi ie s employing single s a is ical o ope a ional esea ch me hods include linea and mul iple disc imina e analysis (DA), logis ic eg ession (LR), p obi analysis, linea and quad a ic p og amming and da a en elopmen analysis. Classi ie s using machine lea ning me hods such as neu al ne wo ks (NN), suppo ec o machine (SVM), decision ees (DT), gene ic and e olu iona y algo i hms (GA), and Bayesian Model Ensembles (BME) ha e also been in es iga ed (Zhao e al., 2017; Asho eh & B a o, 2019, 2021a,b; B a o e al., 2021; B a o & Ayuso, 2020, 2021). Olson, Delen & Meng (2012) ca y ou an ex ensi e compa ison be ween Decision T ees and o he con en ional da a mining echniques such as Neu al Ne wo ks and Suppo Vec o Machines in an a emp o de elop mo e anspa en and anspo able decision suppo ools o p edic ing bank up cy and simila compa a i e wo k is done by Huang e al. (2004). Di ick e al. (2017) employ su i al analysis echniques o achie e highe le els o ideli y in calcula ing a ime ho izon o when a gi en loan will de aul , hus acili a ing a e u n- ocused pe spec i e (as opposed o bina y classi ie s which a e ubiqui ous in he ield). A simila app oach is aken by Byanjanka (2017) o he online P2P lending ma ke whe e p omising esul s a e achie ed o p edic ing he su i al p obabili ies o bo owe s a di e en imes. Đu o ić (2017) analyses he p obabili y o de aul in a da ase o issued loans om online P2P lending pla o m Lending Club, by employing a su i al analysis model o analyze he p obabili y o de aul o loan e m leng h and pu pose. Findings yield ha he e a e iskie loan pu pose classes bu also asse ha a longe e m is associa ed wi h a highe likelihood o de aul . Tan e al. (2018) s udy a simila da ase and p opose an app oach elying on di e en ia ing loans based on nega i e ou comes (ei he “cha ge-o ” o “p e-paymen ”) and, le e aging on deep neu al ne wo ks, ou pe o ming he isk analysis capabili ies o adi ional isk assessmen capabili ies o adi ional su i al analysis echniques. Mezei, Byanjanka & Heikkilä (2018) use linguis ic da a ans o ma ion as a disc e iza ion s ep in adi ional supe ised lea ning algo i hms o success ully imp o e hei classi ica ion pe o mance. Byanjanka , Heikkilä & Mezei (2015) de elop a c edi sco ing model using A i icial Neu al Ne wo ks (ANNs) o e alua e c edi isk in P2P lending, achie ing sa is ac o y esul s. Mo e comp ehensi e wo ks ackle a a ie y o op ions o de e mining isk. Jin & Zhu (2015) de elop a mo e comp ehensi e amewo k o app oaching such da ase s and combine an in oduc o y s ep o iden i ying he de e minan s o isk h ough Random Fo es (RF) and impo ance analysis echniques o hen se e a mul i ude o da a mining algo i hms o compa ison. They conclude ha SVMs ou pe o m o he me hods, al hough by a ma ginal amoun . Ca michael (2014) uses a disc e e- ime haza d model (dynamic logis ic eg ession) o es ima e he p obabili y o de aul o P2P loans. By using a mix o bo h ha d and so in o ma ion, he model can achie e much be e esul s when compa ed o he su eyed pla o m’s c edi g ading sys em. 17 3 METHODOLOGY 3.1 THE MODEL - LOGISTIC REGRESSION In his wo k, we employ a logis ic eg ession model o assess he ac o s ha de e mine he loan pe o mance o P2P loans. Logis ic eg essions model he p obabili y o an e en occu ing depending on he alues o a se o independen a iables, p edic ing hei e ec on a bina y a iable esponse and classi ying obse a ions by es ima ing he p obabili y ha an ou come has in a pa icula ca ego y. This s a is ical echnique has been he mos widely used in c edi sco ing applica ions (Abdou & Poin on, 2011) and is conside ed he indus y s anda d o de eloping c edi sco ing models (Ala’ aj & Abbod, 2015; Lessmann, Baesens, Seow, & Thomas, 2015). Fo his s udy, loan s a us is ou dependen a iable wi h he ollowing bina y ou come: 0 (ze o) i he loan s a us is “Fully Paid”, meaning ha bo owe was able o epay he loan and 1 (one) i he loan s a us is “De aul ”, meaning ha he bo owe is unable o ully pay he loan and de aul 2 . In his sec ion, he wo k o Hosme J e al. (2013) was closely ollowed. Conside ing he abo e, he p obabili y o a loan de aul (condi ional p obabili y) gi en he independen a iable can be w i en as !" ( $!=1|(! ) =*((!) , + whe e he bo owe , +will de aul gi en he in o ma ion (! . I his p obabili y is g ea e han 50%, he model p edic s ha he bo owe belongs o he “De aul ” class, o he wise he model p edic s ha he bo owe ins ead belongs o he “Fully Paid” class. The i s s ep o de e mine his p obabili y is o es ima e a linea eg ession+ unc ion: !(#!) = &' "($") &'"($")= ((+ (&#!& + ⋯ + ()#!) + = 1, 2, … , '0 (1)0 whe e -" is a cons an , -# is a ec o o eg ession coe icien s and ( . ! is a ec o o independen a iables. Thus, #!) is he alue o he a iable Χ!# o he ,$% bo owe , wi h 0 =1,…,3 ( 3 is he numbe o independen a iables). Howe e , in a linea eg ession unc ion, he p obabili ies can be less han 0 and g ea e han 1, hence he model needs o es ic he p obabili y be ween 0 and 1 by including a non-linea unc ion (logi ) on o equa ion (1) ans o ming in o logis ic eg ession unc ion (equa ion 2): 1(#!) = *#(%") &+*#(%") 0 + = 1, 2, … , ' 0 (2)0 Ha ing he model speci ied, he nex s ep is o es ima e he coe icien s -" and -# . The me hod used o i he non-linea model is he Maximum Likelihood Me hod (MLM). This me hod gene a es alues o he unknown coe icien s ha maximize he p obabili y o ob aining he obse ed se o da a. The likelihood unc ion is shown in equa ion (3). 2 Please e e o he nex chap e , Da a, o a u he enligh enmen on he a iable loan s a us 18 4(5)=61(#,)-!71 − 1(#,)81−-! & !'( + + (3)+ I he a iable $! akes he alue 1 ( he bo owe de aul ), he likelihood is *! ; i no , he likelihood is equal o 1−*! . To simpli y, ma hema ically, he op imiza ion o he likelihood unc ion, he loga i hm o he equa ion is no mally maximized. The log-likelihood is de ined as: :(5)=ln[4(5)]?3,&'71(#,)8+@1 − 3,A&'71 − 1(#,)8 & !'( + + (4)+ + The e o e, o ind he alue o 5 ha maximizes he compu a ion : ( 5 ) we can isola e he cases -" and -# , se ing he equa ions equal o ze o and sol ing he p oblem o ob ain he desi ed esul . Due o he nonlinea i y o logis ic eg ession equa ions, so wa e-based compu a ion me hods mus be le e aged o compu e hem, add essing he i e a i e calcula ions equi ed. + 3.2 PERFORMANCE MEASUREMENTS OF THE MODEL To assess he p edic i e capaci y o ou model, we eco d and compa e i s p edic ions o he ac ual, known, obse a ions. A T ue Posi i e (TP) occu s whene e a posi i e p edic ion (ou come equal o “1”, o “de aul ”) made by he model coincides wi h he obse a ion, whe eas a False Posi i e (FP) is eco ded when he model’s posi i e p edic ion is w ongly classi ied and con adic s he obse a ion. T ue Nega i es (TN) and False Nega i es (FN) deno e a simila logic, bu o nega i e p edic ions (ou come equal o “0”, o “ ully paid”). Table 1 - Con usion Ma ix Obse ed Posi i e Nega i e P edic ed Posi i e TP FP Nega i e FN TN Sou ce: Au ho ’s p oduc ion. To e i y he p edic i e capaci y o ou logis ic eg ession model, i is necessa y o unde s and i s disc imina ion unc ion. Fo his, and conside ing he abo e con usion ma ix, we can calcula e he ollowing pe o mance measu es: Accu acy measu es he o e all a io o co ec ly p edic ed obse a ions and can gi e an o e all iew o he model's pe o mance. Howe e , looking a his measu e alone can build a alse pe cep ion o he model due o i s na ow ocus. 19 BCCD"EC$ =FG+F! FG+F!+IG+I!+ + (5)+ Sensi i i y co esponds o he p obabili y o he co ec classi ica ion o he e en o in e es o occu , i.e., how well he model co ec ly classi ies ue posi i es. JKLM,N,O,N$ =F! F!+IG+ + (6)+ Speci ici y is de ined as he p opo ion o obse a ions co ec ly p edic ed o belong o he nega i e class. JPKC,Q,C,N$ =FG FG+I!+ + (7)+ Al e na i e me hods o e alua e he pe o mance o a classi ica ion model wi h bina y classes include he Recei e Ope a ing Cha ac e is ic (ROC) cu e and he A ea Unde he Cu e (AUC) (Wendle & G ö up, 2016). The ROC shows he alues o which he e is g ea e op imiza ion o Sensi i i y as a unc ion o Speci ici y, i.e., plo s he ue posi i e a e (TPR o Sensi i i y) o a disc e e classi ie agains he alse posi i e a e (FPR o 1-Speci ici y). The AUC, in u n, being a unc ion o bo h he TPR and he FPR, p o ides a measu e o he model’s abili y o di e en ia e be ween classes, o i s disc imina ion powe – how e ec i ely i can p edic nega i e and posi i e ou comes co ec ly. As a baseline we can unde s and how he op imal alue o AUC would be 1 – he cu e would pass by poin (0,1), which would hen maximize AUC and gi e he op imal esul o 100% TPR and 0% FPR. Con a ily o his, he baseline unc ion d awn o * ( (! ) in he ROC cu e plo p o ides he lowe bound o a model wi hou classi ica ion powe and a calcula ed alue o 0.5 o AUC. 20 4 DATA This chap e is di ided in o 4 sec ions. In he i s sec ion, “Da a Desc ip ion”, we p o ide a b ie , high- le el desc ip ion o he da a collec ed. In he second pa , “Da a Re inemen ”, we p oceed o desc ibe he s eps aken o clean up he da ase and elabo a e on how he da a e inemen was accomplished. A e ha ea men o he s a ing da ase , we conduc an explo a o y analysis o he new, e ined, da ase in he hi d sec ion “Explo a o y Da a Analysis”. The e we summa ize he key obse a ions on a iables o he da ase , iden i ying ou lie s and handling missing alues. A he end o his sec ion, we p o ide he desc ip i e s a is ics and co ela ion be ween he selec ed a iables. 4.1 DATA DESCRIPTION The da a unde analysis we e e ie ed om he Lending Club websi e 3 , he la ges P2P lending pla o m in he Uni ed S a es. The da a collec ed con ains in o ma ion on mo e han 2,200,000 loans issued h ough he pla o m om 2007 o 2019. Al hough he la es obse a ions in he da ase a e om Ma ch 2019, i con ains only loans issued un il Decembe 2018. Figu e 2 shows he numbe o loans issued h oughou he yea s and in i , we can obse e a p ominen g ow h end in he o al numbe o issued loans un il 2015, whe e he numbe o loans app oxima ely doubles yea - o-yea . F om 2015 onwa ds he e is s ill an inc ease, bu i is subdued when compa ed o he p e ious ime ame. Figu e 2 - Numbe o loans issued by yea Sou ce: Au ho ’s p oduc ion. The o iginal da ase is composed o 151 a iables co e ing de ails ac oss 3 main opics, which a e ocused on: bo owe de ails, loan cha ac e is ics and inal loan s a us (whe e a ailable). Bo owe in o ma ion includes da a poin s such as occupa ion, employmen de ails, c edi a ing and o he addi ional in o ma ion which p o ides Lending Club use s wi h an o e iew o who he bo owe is 3 Downloaded om h ps://www.lendingclub.com/in o/download-da a.ac ion 21 and wha hei cu en inancial/labo s a us is. Loan cha ac e is ics include de ails such as he amoun o he loan, in e es a e, mon hly ins allmen , pu pose o he loan, and o he loan-speci ic elemen s, ha cla i y how he loan is o be epaid. Finally, he in o ma ion p o ided on he inal loan s a us in o ms o he la es s a us o he loan, a he ime o obse a ion (i.e. he loan can ei he be paid o , de aul ed, o e due o cu en ) and p o ides addi ional con ex , depending on ha same s a us – e.g. is he e a paymen plan o he de aul ed loan. The ull desc ip ion o all 151 a iables can be ound in Appendix A. Ou o 151 a iables, we ha e selec ed 22 independen a iables, as no all he a iables a e ele an o his analysis. The selec ion o he a iables in he s udy is explained u he in he ollowing sub-sec ions. 4.2 DATA REFINEMENT In his sec ion, we desc ibe all he s eps in ol ed in he ea men o he da ase and how we ha e p oceeded o e ine he da a he ein. The i s s ep in ou ea men o he da ase consis s o na owing down he ime ame unde s udy. We p oceed o il e ou he loans ha ha e no eached ma u i y ye , o gua an ee consis ency be ween he da a poin s a ailable. Wi h his, we ha e emo ed all loans wi h a 36-mon h epaymen e m ha we e issued a e Feb ua y 2016 and all loans wi h a 60-mon h epaymen e m ha we e issued a e Feb ua y 2014. The loans issued in 2007 we e also emo ed as he bo owe in o ma ion ini ially eques ed was di e en om he ollowing yea s. As p e iously men ioned, no all he a iables a e ele an o ou s udy. Ou goal wi hin his sec ion is o iden i y p edic o s o de aul so ha a he ime o he loan eques , he pla o m can use hose a iables as a e e ence o decide on whe he o app o e o ejec he loan. Due o his, any a iables ha a e no a ailable a he ime o he loan applica ion canno be used as p edic o s o c edi app o al. Fo his eason, we emo ed a iables such as nex _pymn _d, ha dship_ lag, pymn _plan among o he s ela ed o hese. We lea e ou some a iables because hey ep esen he same da a as o he s, such as unded_amoun and unded_amoun _in ha ep esen he same as loan_amn . O he a iables a e also deemed o no be ele an as hey don’ ha e any impac on he bo owe ’s likelihood o de aul , such as id, membe _id and policy_code, and gi en his, hey a e also emo ed. We also p oceeded o emo e da ase columns such as emp_ i le and u l, because hey con ain oo many unique alues, making hem di icul o ca ego ize and analyze, so we op o lea e hem ou o he s udy. Ano he eason o dele e columns is he high equency o missing da a in some a iables. We ha e decided o lea e ou all he a iables wi h mo e han i e pe cen (5%) o missing alues. The e o e, we ha e le ou a iables such as m hs_since_las _ eco d and m hs_since_las _delinq. A u he e inemen is conduc ed on he emp_leng h a iable, due o i ha ing a subs an ial olume o missing alues. To p e en disca ding hese da a poin s (which amoun o app oxima ely 6% o he o al da ase ), we emploied a ‘binning” echnique and g oup all missing alues in o a sepa a e g oup labeled “missing”. Figu e 3 illus a es he dis ibu ion o he numbe o obse a ions by he leng h o employmen . 22 Figu e 3 - Numbe o obse a ions by employmen leng h Sou ce: Au ho ’s p oduc ion. I can be obse ed in Figu e 3 ha he majo i y o he ca ego ies ha e a subs an ially educed numbe o obse a ions when compa ed wi h he ca ego y “10+ yea s”. We chose o agg ega e some ca ego ies oge he , esul ing in wo new ca ego ies (“0 - 4 yea s” and “5 - 9 yea s”). This s ep was execu ed o simpli y he ca ego ies a ailable in a iable emp_leng h. Table 2 below shows he new dis ibu ion o he a iable emp_leng h. Table 2 - Summa y o he employmen leng h a iable Missing 0 - 4 yea s 5 - 9 yea s 10+ yea s 41,645 278,428 183,607 231,919 Sou ce: Au ho ’s p oduc ion. The only o he a iables wi h missing alues we e e ol_u il and pub_ ec_bank up cies, wi h 0.06% and 0.11% o missing alues, espec i ely. As hese pe cen ages ep esen a o al o 1,208 obse a ions we decided o emo e hese missing alues. As p e iously s a ed, loan s a us is he dependen a iable o his s udy and i is di ided in o six ca ego ies desc ibed in Table 3. To adap he a iable o ou model, we ha e o ans o m i in o a a iable con aining a bina y ou come/ alue, 0 o 1. When he loan_s a us assumes he alue o one (1) i ep esen s he ailu e o paymen , meaning ha he bo owe is unable o ully pay he loan and e en ually de aul s. On he o he hand, when loan_s a us assumes he alue ze o (0), his means ha he bo owe had paid he loan, ul illing he es ablished con ac . 23 Table 3 - Numbe o loans by loan s a us Ini ial Da a Se Dis ibu ion Final Da a Se Dis ibu ion Loan S a us # o Loans Loan S a us # o Loans Cha ged O 109,258 1 - De aul 109,797 Cu en 433 0 - Fully Paid 623,770 Does no mee he c edi policy. S a us: Cha ged O 539 Does no mee he c edi policy. S a us: Fully Paid 1,521 Fully Paid 622,249 In G ace Pe iod 40 La e (16-30 days) 34 La e (31-120 days) 317 To al 734,391 To al 733,567 Sou ce: Au ho ’s p oduc ion. To p oceed wi h he ans o ma ion o he a iable, we ha e emo ed all he loans wi h “Cu en ” s a us as he bo owe s a e s ill making mon hly paymen s and we don’ ye know he inal s a us. Loans wi h he “In G ace Pe iod” s a us ep esen he loans ha a e a he mos 15 days wi h a la e paymen and loan s a uses wi h names “La e (16-30 days)” and “La e (31-120 days)” ep esen he loans ha e a delayed ins allmen be ween 16-30 and 31-120 days, espec i ely. While hese loans a e in some way in de aul , we il e ed hem ou in his s udy as he p obabili y o hem being paid back is signi ican (26% o 72%) compa ed wi h he “De aul ” s a us (11%) (Lending Club, 2019a). We hen p oceeded o agg ega e he s a us “Cha ged O ” wi h “Does no mee he c edi policy. S a us: Cha ged O ” in o a newly c ea ed s a us labeled as “De aul ”, since bo h co espond o he same ou come. The s a us “Does no mee he c edi policy. S a us: Fully Paid” is also combined in o he “Fully Paid” ca ego y. Table 3 shows he ini ial dis ibu ion o he numbe o loans by s a us and he inal dis ibu ion a e he adjus men s. The loan desc ip ion p o ided by he bo owe on why he loan is needed is con ained in he a iable desc and i is op ional in o ma ion ha mos o he bo owe s lea e blank. The a iable by i sel is di icul o analyze since his is a ee-s yle ield. To e ie e alue ou o his a iable, we c ea ed a dummy a iable, w i ing_skills, ha is de i ed om he o iginal use inpu . I he bo owe has le a desc ip ion on his ield, he o she will ha e au oma ically 10 poin s. Then, we check i he e a e any spelling mis akes on he desc ip ion - i de ec ed, 2 poin s will be deduc ed om he ini ial 10. We also check i he ex has no end ma ks, i.e., i i is missing ending punc ua ion - i de ec ed, ano he 2 poin s will be deduc ed. The possible alues o w i ing_skills a e 0, i he bo owe hasn’ w i en any desc ip ion, 6 i he e is any wo d misspelled and any missing punc ua ion, 8 i he e is any wo d misspelled o any missing punc ua ion, and 10 i he e is a desc ip ion wi h no spelling mis akes and wi h co ec punc ua ion. Al hough he desc a iable is only illed in by app oxima ely 17% o bo owe s, ea lie wo ks ha e indica ed ha his a iable may be a good p edic o o de aul (Ca michael, 2014). We employ his 24 composi ion o ex ual analysis me hods o de i e nume ic da a which is easie o analyze and o use wi hin he con ex o ou p edic ion model. Las ly, he da e o he ea lies c edi line opened by he bo owe (mon h and yea ) is shown in ea lies _c _line. We ha e c ea ed a dummy a iable, c edi _his o y, whe e we ha e sub ac ed he loan’s issue da e by he da e o he ea lies c edi line. We now ha e he numbe o yea s o c edi his o y conside ing he da e when he bo owe ’s ea lies epo ed c edi line was opened. The a o emen ioned e inemen s eps esul in he educ ion o a iables om a s a ing base o 151 o a o al o 25. In he sub-sec ion ha ollows we p oceed o analyze he esul ing da ase and conduc any addi ional e inemen ha may be deemed necessa y, be o e building ou model. 4.3 EXPLORATORY DATA ANALYSIS To u he e ine ou da ase and o p epa e i o usage in building ou p edic ion model, we ha e conduc ed an explo a o y analysis o he p o ided da a. We begin his explo a o y da a analysis by in es iga ing he ou lie s ha may exis on he da ase . Ou i s inding ocuses on he a iable annual_inc. Table 4 p o ides a summa y o his a iable and a i s glance, we can de ec some alues ha seem un ealis ic o his da a. The maximum annual income o $9,000,000 compa ed wi h he median annual income o $62,000 and he 3 d qua ile o $89,000 sugges s ha he e a e some ou lie s. Figu e 4 shows he dis ibu ion o bo owe s’ annual income. Obse ing Table 4 along wi h Figu e 4, we can con i m he e idence o ou lie s. Table 4 - Summa y o he annual income a iable Min. 1s Qu. Median Mean 3 d Qu. Max. 0 45,000 62,000 73,689 89,000 9,000,000 Sou ce: Au ho ’s p oduc ion. A deepe analysis o he emaining in o ma ion p o ided o hese obse a ions shows ha he majo i y o he bo owe s a e eques ing small loan amoun s o deb consolida ion. Taking all hese de ails in o accoun , we conside hese eco ds o be he esul o human e o and, he e o e, we ha e limi ed he highes alue o annual income o $1,000,000. In o al, he e we e 139 obse a ions emo ed. 25 Figu e 4 - Dis ibu ion o annual income Sou ce: Au ho ’s p oduc ion. Ano he inding conce ns he d i a iable (deb - o-income a io), which ep esen s he a io o he bo owe ’s o al mon hly deb paymen s o e hei mon hly income. Figu e 5 displays he boxplo o he a iable d i. Figu e 5 - D i ou lie s Sou ce: Au ho ’s p oduc ion. In obse ing Figu e 5 i becomes no iceable ha he majo i y o loans a e concen a ed in he ange o alues be ween 10 and 25 and he e is e idence o se e al ou lie s. We belie e ha hese ou lie s a e caused by an e o as he da a p o ided would indica e ha some bo owe s ha e signi ican ly mo e ou s anding deb han hei cu en income – we also obse ed he alue o d i o be inconsis en wi h some o he o he da a poin s p o ided, o some bo owe s – hus, we ha e emo ed all he ou lie s. Compa ed wi h o he s udies, o ins ance, Ca michael (2014), Emek e e al. (2015) and Möllenkamp (2017), we can see ha a e he emo al o he ou lie s, he maximum alue o his a iable is as expec ed. The exclusion o d i’s ou lie s dec eases ou da ase by 484 obse a ions. 32 The same happens wi h he in e es a e. This a iable is also highly signi ican on he p edic ion o he de aul ha wi h he inc ease o he a e, he p obabili y o de aul also inc eases. This inding is consis en wi h he indings in Figu e 6 ep esen ed in he p e ious chap e . Loan amoun and e m a e o he a iables o be highly signi ican in p edic ing he ailu e o he loan paymen . The posi i e sign on hese a iables sugges s ha bo owe s wi h a la ge loan amoun and a longe pe iod o loan paymen a e mo e likely o de aul . E e y hing else equal, his p o es ha bo owe s who eques la ge loan amoun s on a longe - e m and he e o e will ha e la ge ins allmen s o a longe pe iod o paymen a e mo e likely o ail he paymen s. The a iable home owne ship shows ha a bo owe paying en o owning a house is mo e likely o de aul han a bo owe ha pays a mo gage. These esul s a e in line wi h he da a p e iously e ealed on he explo a o y analysis whe e hese wo ca ego ies ha e a highe de aul a e han he la e one. Fu he o his, he wo ca ego ies ha ha e some kind o e i ica ion on he a iable e i ica ion s a us a e also highly signi ican p edic o s. The esul s a e in line wi h he da a examined in he p e ious chap e . This may sugges ha Lending Club e i ies he bo owe s ha hey belie e o ha e a highe isk o de aul . Annual income, FICO and d i a e also highly signi ican a iables. The nega i e sign on he wo i s a iables implies ha a bo owe wi h a highe income and a g ea e FICO sco e is less likely o de aul . Con a ily, he las o hese a iables wi h a posi i e sign indica es ha bo owe s wi h highe amoun s o deb s uggle mo e o epay hei loans. The bo owe ’s pu pose o eques ing a loan has 14 ca ego ies. O hose 14, 10 ca ego ies a e signi ican a di e en le els. Bo owe s needing a loan o help hem inance hei “educa ion”, “small business” o “medical expenses” a e iskie bo owe s han hose who a e inancing hei “wedding”, “ aca ion” o epaying hei “c edi ca d” deb . Ac ually, he ca ego y “wedding” egis e s a nega i e co ela ion wi h he likelihood o de aul e ealing o be he only pu pose ha a ec s nega i ely he p obabili y o de aul wi h an OR o 0.824. On he a iable employmen leng h, al hough he ca ego y 5-9 yea s is signi ican (a 5% le el), only he ca ego y “missing” is highly signi ican a he 0.1% le el. Also, he la e has a coe icien wi h a posi i e sign and OR o 1.516, ha ing a big impac on he likelihood o de aul . The ca ego y 10+ yea s, despi e no being signi ican , is he only one om his a iable wi h a nega i e sign, ha ing he opposi e e ec on he p obabili y o de aul . These esul s a e aligned wi h he da a in Table 12 o Appendix C. A possible jus i ica ion o he inc eased likelihood o a bo owe de aul ing, when egis e ed unde he “missing” ca ego y, may be due o him being unemployed and, hence, being mo e likely o de aul . Rega ding he a iables ela ed o he c edi his o y o he bo owe , inqui ies las 6 mon hs, open accoun s, e ol ing balance, e ol ing line u iliza ion a e and c edi his o y a e also highly signi ican . This is in line wi h ou expec a ions, inqui ies las 6 mon hs and open accoun s show he bo owe s lack epaymen o a loan, inc easing he p obabili y o de aul . On he o he hand, e ol ing balance, e ol ing line u iliza ion a e, and c edi his o y show a nega i e ela ionship wi h he likelihood o de aul o a bo owe . 33 Su p isingly, he a iable cha ged o wi hin 12 mon hs has a nega i e impac on he p obabili y o de aul o he bo owe , i.e., he bo owe wi h mo e cha ge-o s (on o he loans) in he las 12 mon hs is mo e likely o ully pay back he loan. The a iable w i ing skills is highly signi ican and a ec s he p obabili y o de aul nega i ely. When he alue o w i ing skills inc eases, he p obabili y o he bo owe paying also inc eases. Rega ding public eco ds, he a iable seems no o ha e any signi icance on he model in es iga ed, howe e has an odds a io o 1.011, which is highe han o he a iables wi h a highe le el o signi icance. 5.2 MODEL PERFORMANCE The p edic i e capaci y o he model was es ed by le e aging he emaining 33% o he sample da ase which we e no used o building ou logis ic eg ession model (i.e. used in he “ aining” se ). We hen p oceeded o use his “ es ing” se o e ie e a se ies o pe o mance me ics o analyze he model, as desc ibed in sec ion 3.2. Table 7 p o ides he summa y o he ou comes in a con ingency able. Table 7 - Con usion Ma ix Obse ed Posi i e Nega i e P edic ed Posi i e 207745 79 Nega i e 36292 64 Sou ce: Au ho ’s p oduc ion. Based on he alues o he abo e able, we calcula ed he ollowing pe o mance measu es. Table 8 - Model Pe o mance Measu es Accu acy 0.851 Sensi i i y 0.851 Speci ici y 0.448 Sou ce: Au ho ’s p oduc ion. The accu acy alue o he es da ase in he s udy has a e y good ou come. Ne e heless, we ha e in mind ha his esul is dependen on he manual spli o he da a ha we did ea lie and ha his measu e alone is no a eliable indica o o he model’s pe o mance. Conce ning he o he wo measu es, he model pe o ms be e in dis inguishing bo owe s ha de aul ed han bo owe s ha ully paid, as i yields be e esul s o sensi i i y han speci ici y. Figu e 8 shows he ROC cu e o ou model. As we can obse e, he cu e is e y close o he baseline (black line ha ep esen s he balance TPR=FPR). 34 Figu e 8 - ROC Cu e Sou ce: Au ho ’s p oduc ion. By e iewing he AUC indica o , i s esul e eals a alue o 0.6835, showing ha , al hough ou model can accu a ely p edic loans ha will de aul , i is a less capable a p edic ing loans ha will no do so. This ep esen s a lack o abili y o di e en ia e be ween he 2 classes o ou dependen a iable – “de aul ” VS “ ully paid”. We can conclude om he esul s and ela ed pe o mance measu es ha he model’s classi ica ion powe is a he bes “a e age”. We belie e ha his lack o quali y may be due o he imbalanced classi ica ion in ou dependen a iable as i is a common limi a ion wi hin loan classi ica ions and c edi isk modeling. As pe Singh, Tsai, & Ramiah (2014), his imbalance nega i ely impac s algo i hms such as Logis ic Reg ession ha op imizes ac oss he en i e aining se . 35 6 CONCLUSIONS In an inc easingly inancialized wo ld, whe e new inancial echnologies and se ice models keep on dis up ing he es ablished and adi ional p ac ices and indus ies, pee - o-pee (P2P) lending seeks o democ a ize access o c edi o a wide ange o bo owe s (including hose deemed by con en ional inance o be non-c edi wo hy) and also o unsecu ed consume deb , o a wide ange o in es o s. In doing so, i is b idging a demand and a supply gap which cu en ly exis s in his la gely un apped ma ke . Lending Club, being one o he pionee s and bigges P2P lending pla o ms, p o ides a good objec o s udy and aluable insigh in o he inne wo kings o hese pla o ms and hei abili y o adequa ely p edic he isk o c edi de aul . This las unc ion u ns Lending Club (and o he analogous pla o ms) in o mo e han jus a media o connec ing supply and demand, assigning i he esponsibili y o cu a ing and a es ing bo owe in o ma ion and isk. In he beginning o his wo k, we se ou o s udy a da ase o loan obse a ions om he Lending Club pla o m, con aining in o ma ion o loans issued be ween 2007 and 2018. Wi h i we endea o ed o u he he cu en unde s anding and es ablished knowledge o wha a e he de e minan s o loan de aul and wha pe o mance we can expec om a p edic ion model buil a op his da a. In o de o achie e his, we i s conduc ed a cleaning and e inemen o he da a se , elimina ing inco ec and missing da a, whe e equi ed. This p ocess was ollowed by an explo a o y analysis o he ou pu ed da a, o gain an o e all unde s anding o he exis ing a iables, hei co ela ion and d i ing he selec ion o he inal a iables o be used o ou s udy o de e minan s o de aul . To u he ou analysis, we buil a logis ic eg ession model o analyze he da ase using he s epwise backwa d echnique in o de o a i e a a inal, model-de i ed, se o independen a iables ha we e he mos signi ican o p edic ing loan de aul . Ou analysis aims a es ablishing a li ecycle o e iewing and unde s anding bo owe de aul based on ha d da a eadily a ailable o he in es o (o o P2P lending pla o ms). Unlike p io wo k, which ocused on de eloping p edic i e models o iden i ying he de e minan s o de aul , ou wo k lays ou and end- o-end ounda ion o unde s anding he cha ac e is ics o bo owe s loans wi hin a pla o m (o se o loan obse a ions), iden i ying de e minan s o de aul , building a p edic i e model and analyzing i s e icacy in p edic ing loan de aul . Ou esul s seem o be consis en wi h exis ing li e a u e (Ca michael, 2014; Emek e e al., 2015; Polena & Regne , 2018; Se ano-Cinca e al., 2015) and sugges ha he a iables which bes explain (o p edic ) he bo owe s likelihood o de aul a e g ade, loan amoun , in e es a e, e m, annual income, home owne ship, deb - o-income a io, FICO sco e, inqui ies in las 6 mon hs, numbe o open accoun s and c edi his o y. The mos ele an independen a iable is he g ade o a bo owe (as assigned by Lending Club) which, ha ing he highes odds a io, p esen s he bes p edic o o loan de aul . The highe he g ade o a gi en loan (“A” being he highes g ade and “G” he lowes , in alphabe ical o de ), he highe he likelihood ha i will be ully epaid. 36 Finally, al hough ou model has shown a high accu acy a e o 0.8511, he AUC analysis p oduced a esul o 0.6835, meaning ha he o e all pe o mance o ou model, as ep esen ed by i s abili y o dis inguish be ween classes (i.e. bo owe s ha will de aul VS bo owe s ha will ully pay back a loan) is a om excellen and, a bes , medioc e. In summa y, al hough ou model has been able o e ec i ely p edic de aul ed loans, i s esul s ha e yielded a signi ican amoun o “ alse nega i es”, i.e. ully paid loans ha we e p edic ed o de aul . We ex apola e om his ha , al hough oppo uni ies may be missed o und good quali y loans, i is bes o e on he side o cau ion when dealing wi h unsecu ed consume deb . 6.1 LIMITATIONS AND FUTURE WORK Th oughou his wo k, we iden i ied se e al limi a ions and cons ain s ha may ha e hinde ed ou in es iga ion and i s subsequen esul s. We p esen hose in his sec ion and also a emp o lay ou he ounda ion o u u e wo k o build upon he esul s p o ided he ein. Du ing he ini ial analysis and clean-up/ e inemen o ou da ase , we we e o ced o emo e a subs an ial numbe o a iables and loan obse a ions due o missing da a o inco ec ly illed ou o ms. This app oach, al hough consis en wi h es ablished da a analysis p ac ices, may ha e caused he loss o po en ially aluable da a poin s o he s udy. We hypo hesize ha he cause o his p oblem, wi h a special ocus on inco ec ly p o ided da a, is likely ied o he ac ha hese pla o ms allow o he manual in oduc ion o da a by use s. Such pla o ms also lack he manpowe o capaci y o handle and p ope ly e 100% o he da a which hey ecei e. Wi h ega ds o he esul s ob ained wi h he de eloped logis ic eg ession model, we es ima e ha he medioc e p edic i e pe o mance, which we achie ed, is a p oduc o he da ase being imbalanced. This is a common issue in c edi da ase s as hese will la gely con ain a disp opo iona e amoun o non-de aul ed loans (85% o non-de aul ed loans in ou da ase ) and one ha nega i ely impac s he p edic i e capaci y o logis ic eg ession models (Singh e al., 2014). 37 7 BIBLIOGRAPHY Abdou, H. A., & Poin on, J. (2011). C edi sco ing, s a is ical echniques and e alua ion c i e ia: a e iew o he li e a u e. In elligen Sys ems in Accoun ing, Finance and Managemen , 18(2–3), 59–88. Ala’ aj, M., & Abbod, M. (2015). A sys ema ic c edi sco ing model based on he e ogeneous classi ie ensembles. 2015 In e na ional Symposium on Inno a ions in In elligen SysTems and Applica ions (INISTA), 1–7. Asho eh, A. & B a o, J. M. (2019). A non-pa ame ic based compu a ionally e icien app oach o c edi sco ing. A as da Con e encia da Associacao Po uguesa de Sis emas de In o macao 2019 [CAPSI 2019 - 19 h Con e ence o he Po uguese Associa ion o In o ma ion Sys ems, P oceedings. 4]. h ps://aisel.aisne .o g/capsi2019/4. Asho eh, A., & B a o, J. M. (2021a). A Conse a i e App oach o Online C edi Sco ing. Expe Sys ems Wi h Applica ions, Volume 176, p. 1-16, 114835. h ps://doi.o g/10.1016/j.eswa.2021.114835 Asho eh, A. & B a o, J. M. (2021b). Spa k Code: A No el Conse a i e App oach o Online C edi Sco ing [Sou ce Code]. h ps://doi.o g/10.24433/CO.1963899. 1. Associa ed Publica ion: “A Conse a i e App oach o Online C edi Sco ing”, Expe Sys ems wi h Applica ions, h ps://doi.o g/10.1016/j.eswa.2021.114835 Ba h, C. (2012). Looking Fo 10% Yields? Go Online Fo Pee To Pee Lending. Fo bes. Re ie ed om h ps://www. o bes.com/si es/ch isba h/2012/06/06/looking- o -10-yields-go-online- o -pee - o-pee -lending/#6396ea a3a8 Be ge , S. C., & Gleisne , F. (2009). Eme gence o inancial in e media ies in elec onic ma ke s: The case o online P2P lending. BuR Business Resea ch Jou nal, 2(1). Boo , A. W. A., & Thako , A. V. (1997). Financial sys em a chi ec u e. The Re iew o Financial S udies, 10(3), 693–733. B adley, Ch is ine; Bu house, Susan; G a on, Hea he ; Mille , R.-A. (2009). Al e na i e Financial Se ices: A P ime . FDIC Qua e ly, 3(1), 39–47. Re ie ed om h ps://www. dic.go /bank/analy ical/qua e ly/2009- ol3-1/ dic140-qua e ly ol3no1-a s- inal.pd B a o, J. M. (2020). Longe i y-Linked Li e Annui ies: A Bayesian Model Ensemble P icing App oach. CAPSI 2020 P oceedings, 29. h ps://aisel.aisne .o g/capsi2020/29. B a o, J. M. (2021). P icing pa icipa ing longe i y-linked li e annui ies: A Bayesian Model Ensemble app oach. Eu opean Ac ua ial Jou nal. h ps://doi.o g/10.1007/s13385-021-00279-w B a o, J. M., Ayuso, M. (2020). Mo ali y and li e expec ancy o ecas s using bayesian model combina ions: An applica ion o he po uguese popula ion. RISTI - Re is a Ibé ica de Sis emas e Tecnologias de In o mação, E40, 128–144. h ps://doi.o g/10.17013/ is i.40.128–145. B a o, J. M., Ayuso, M. (2021). Fo ecas ing he e i emen age: A Bayesian Model Ensemble App oach. Ad ances in In elligen Sys ems and Compu ing, Volume 1365 AIST, 123–135 [2021 Wo ld Con e ence on In o ma ion Sys ems and Technologies, Wo ldCIST 2021] Sp inge , Cham. h ps://doi.o g/10.1007/978-3-030-72657-7_12. B a o, J. M., Ayuso, M., Holzmann, R., Palme , E. (2021). Add essing he Li e Expec ancy Gap in 38 Pension Policy. Insu ance: Ma hema ics and Economics, 99, 200-221. h ps://doi.o g/10.1016/j.insma heco.2021.03.025. Byanjanka , A. (2017). P edic ing c edi isk in Pee - o-Pee lending wi h su i al analysis. 2017 IEEE Symposium Se ies on Compu a ional In elligence (SSCI), 1–8. Byanjanka , A., Heikkilä, M., & Mezei, J. (2015). P edic ing c edi isk in pee - o-pee lending: A neu al ne wo k app oach. 2015 IEEE Symposium Se ies on Compu a ional In elligence, 719–725. Ca michael, D. (2014). Modeling de aul o pee - o-pee loans. A ailable a SSRN 2529240. Chamboko, R., & B a o, J. M. (2016). On he modelling o p ognosis om delinquency o no mal pe o mance on e ail consume loans. Risk Managemen , 18(4), 264–287. h ps://doi.o g/10.1057/s41283-016-0006-4 Chamboko, R., & B a o, J. M. (2019a). F ail y co ela ed de aul on e ail consume loans in Zimbabwe. In e na ional Jou nal o Applied Decision Sciences, 12(3), 257–270. h ps://doi.o g/10.1504/IJADS.2019.100436 Chamboko, R., & B a o, J. M. V. (2019b). Modelling and o ecas ing ecu en eco e y e en s on consume loans. In e na ional Jou nal o Applied Decision Sciences, 12(3), 271–287. h ps://doi.o g/10.1504/IJADS.2019.100440 Chamboko, R., & B a o, J. M. (2020). A Mul i-S a e App oach o Modelling In e media e E en s and Mul iple Mo gage Loan Ou comes. Risks, 8(2), 64. h ps://doi.o g/10.3390/ isks8020064 Cho zempa, M. (2018). Massi e P2P Failu es in China: Unde g ound Banks Going Unde . Pe e son Ins i u e o In e na ional Economics, 21. Re ie ed om h ps://www.piie.com/blogs/china- economic-wa ch/massi e-p2p- ailu es-china-unde g ound-banks-going-unde Co po a e Finance Ins i u e. (n.d.). Pee - o-Pee Lending - O e iew, How I Wo ks, P os & Cons. Re ie ed om h ps://co po a e inanceins i u e.com/ esou ces/knowledge/ inance/pee - o- pee -lending/ Di ick, L., Claeskens, G., & Baesens, B. (2017). Time o de aul in c edi sco ing using su i al analysis: a benchma k s udy. Jou nal o he Ope a ional Resea ch Socie y, 68(6), 652–665. Đu o ić, A. (2017). Es ima ing P obabili y o De aul on Pee o Pee Ma ke – Su i al Analysis App oach. Jou nal o Cen al Banking Theo y and P ac ice, 6(2). Emek e , R., Tu, Y., Ji asakuldech, B., & Lu, M. (2015). E alua ing c edi isk and loan pe o mance in online Pee - o-Pee (P2P) lending. Applied Economics, 47(1), 54–70. h ps://doi.o g/10.1080/00036846.2014.962222 FICO. (2018). An e olu ion in ML inno a ions ha helps bo h lende s and consume s. Re ie ed om h ps://www. ico.com/en/ esou ce-download- ile/6559 Galloway, I., & o he s. (2009). Pee - o-pee lending and communi y de elopmen inance. Communi y In es men s, 21(3), 19–23. Gonzalez, L., & Lou ei o, Y. K. (2014). When can a pho o inc ease c edi ? The impac o lende and bo owe p o iles on online pee - o-pee loans. Jou nal o Beha io al and Expe imen al Finance, 2, 44–58. G eenbaum, S. I., Thako , A. V, & Boo , A. W. A. (2019). Con empo a y inancial in e media ion. Academic P ess. 39 Ha ylchyk, O., Ma io o, C., Rahim, T.-U.-, & Ve die , M. (2016). Wha D i es he Expansion o he Pee - o-Pee Lending? SSRN Elec onic Jou nal. h ps://doi.o g/10.2139/ss n.2841316 Ha ylchyk, O., & Ve die , M. (2018). The inancial in e media ion ole o he P2P lending pla o ms. Compa a i e Economic S udies, 60(1), 115–130. He zens ein, M., And ews, R. L., Dholakia, U. M., & Lyand es, E. (2008). The democ a iza ion o pe sonal consume loans? De e minan s o success in online pee - o-pee lending communi ies. Bos on Uni e si y School o Managemen Resea ch Pape , 14(6), 1–36. He zens ein, M., Dholakia, U. M., & And ews, R. L. (2011). S a egic he ding beha io in pee - o-pee loan auc ions. Jou nal o In e ac i e Ma ke ing, 25(1), 27–36. Hosme J , D. W., Lemeshow, S., & S u di an , R. X. (2013). Applied Logis ic Reg ession (Vol. 398). John Wiley & Sons. Huang, Z., Chen, H., Hsu, C.-J., Chen, W.-H., & Wu, S. (2004). C edi a ing analysis wi h suppo ec o machines and neu al ne wo ks: a ma ke compa a i e s udy. Decision Suppo Sys ems, 37(4), 543–558. Hulme, M. K., & W igh , C. (2006). In e ne based social lending: Pas , p esen and u u e. Social Fu u es Obse a o y, 11, 1–115. Iye , R., Khwaja, A. I., Lu me , E. F. P., & Shue, K. (2009). Sc eening in new c edi ma ke s: Can indi idual lende s in e bo owe c edi wo hiness in pee - o-pee lending? AFA 2011 Den e Mee ings Pape . Jin, Y., & Zhu, Y. (2015). A da a-d i en app oach o p edic de aul isk o loan o online Pee - o-Pee (P2P) lending. 2015 Fi h In e na ional Con e ence on Communica ion Sys ems and Ne wo k Technologies, 609–613. Kä e , B. (2018). Pee - o-Pee Lending - A (Financial S abili y) Risk Pe spec i e. Re iew o Economics, 69(1), 1–25. Keys, B. J., Mukhe jee, T., Se u, A., & Vig, V. (2010). Did secu i iza ion lead o lax sc eening? E idence om subp ime loans. The Qua e ly Jou nal o Economics, 125(1), 307–362. Kla , M. (2008). Online pee - o-pee lending: a lende s’ pe spec i e. P oceedings o he In e na ional Con e ence on E-Lea ning, E-Business, En e p ise In o ma ion Sys ems, and E- Go e nmen , EEE, 371–375. Langage , C. (2019). How Is My C edi Sco e Calcula ed? Re ie ed om In es opedia websi e: h ps://www.in es opedia.com/ask/answe s/05/c edi sco ecalcula ion.asp Lee, E., & Lee, B. (2012). He ding beha io in online P2P lending: An empi ical in es iga ion. Elec onic Comme ce Resea ch and Applica ions, 11(5), 495–503. Lending Club. (2019a). Demand and c edi p o ile. Re ie ed om Lending Club websi e: h ps://www.lendingclub.com/in o/demand-and-c edi -p o ile.ac ion Lending Club. (2019b). How a e loans lis ed and app o ed o in es ing? – LendingClub. Re ie ed om Lending Club websi e: h ps://help.lendingclub.com/hc/en-us/a icles/215466748-How- a e-loans-lis ed-and-app o ed- o -in es ing- Lending Club. (2019c). Ra es & Fees. Re ie ed om Lending Club websi e: h ps://www.lendingclub.com/public/ a es-and- ees.ac ion 40 Lessmann, S., Baesens, B., Seow, H.-V., & Thomas, L. C. (2015). Benchma king s a e-o - he-a classi ica ion algo i hms o c edi sco ing: An upda e o esea ch. Eu opean Jou nal o Ope a ional Resea ch, 247(1), 124–136. Lin, M., P abhala, N. R., & Viswana han, S. (2013). Judging bo owe s by he company hey keep: F iendship ne wo ks and in o ma ion asymme y in online pee - o-pee lending. Managemen Science, 59(1), 17–35. Lus , D. (2017). Analysis o sco ing in pee - o-pee lending. Mezei, J., Byanjanka , A., & Heikkilä, M. (2018). C edi isk e alua ion in pee - o-pee lending wi h linguis ic da a ans o ma ion and supe ised lea ning. Mild, A., Wai z, M., & Wöckl, J. (2015). How low can you go? O e coming he inabili y o lende s o se p ope in e es a es on unsecu ed pee - o-pee lending ma ke s. Jou nal o Business Resea ch, 68(6), 1291–1305. Möllenkamp, N. (2017). De e minan s o Loan Pe o mance in P2P Lending. Uni e si y o Twen e. Nowak, A., Ross, A., & Yencha, C. (2018). Small Business Bo owing and Pee - o-Pee Lending: E idence F om Lending Club. Con empo a y Economic Policy, 36(2), 318–336. h ps://doi.o g/10.1111/COEP.12252 Olson, D. L., Delen, D., & Meng, Y. (2012). Compa a i e analysis o da a mining me hods o bank up cy p edic ion. Decision Suppo Sys ems, 52(2), 464–473. Polena, M., & Regne , T. (2018). De e minan s o bo owe s’ de aul in P2P lending unde conside a ion o he loan isk class. Games, 9(4), 82. P ospe .com. (2016). Wha is he loan e iew p ocess? How long does i ake? – Help is on he way. Re ie ed om P ospe .com websi e: h ps://p ospe .zendesk.com/hc/en- us/a icles/210013753-Wha -is- he-loan- e iew-p ocess-How-long-does-i - ake- Secu i ies and Exchange Commission. (2015). LendingClub Co p Fo m 10-K o Fiscal Yea Ended Decembe 31, 2014. Re ie ed om h ps://www.sec.go /A chi es/edga /da a/1409970/000119312515070385/d851207d10k.h m Se ano-Cinca, C., Gu ie ez-Nie o, B., & López-Palacios, L. (2015). De e minan s o de aul in P2P lending. PloS One, 10(10), e0139427. Singh, S., Tsai, K., & Ramiah, S. (2014). Pee Lending Risk P edic o . h ps://doi.o g/10.13140/2.1.4810.6567 Tan, F., Hou, X., Zhang, J., Wei, Z., Yan, Z., & Weng, S.-C. (2018). A no el isk assessmen scheme and p ac ice o pee - o-pee lending. P oc. ACM SIGKDD Wo kshop Da a Sci. Fin ech. Tapsco , D., & Williams, A. (2007). The new science o sha ing. The BusinessWeek Wikinomics Se ies. T anspa ency Ma ke Resea ch. (2016). Pee - o-Pee Lending Ma ke - Global Indus y Analysis, Size, Sha e, G ow h, T ends and Fo ecas 2016 - 2024. Re ie ed om h ps://www. anspa encyma ke esea ch.com/pee - o-pee -lending-ma ke .h ml Weiss, G. N. F., Pelge , K., & Ho sch, A. (2010). Mi iga ing ad e se selec ion in P2P lending - Empi ical e idence om p ospe . com. A ailable a SSRN 1650774. Wendle , T., & G ö up, S. (2016). Da a mining wi h SPSS modele : heo y, exe cises and solu ions. 41 h ps://doi.o g/10.1007/978-3-319-28709-6 Yum, H., Lee, B., & Chae, M. (2012). F om he wisdom o c owds o my own judgmen in mic o inance h ough online pee - o-pee lending pla o ms. Elec onic Comme ce Resea ch and Applica ions, 11(5), 469–483. Zhao, H., Ge, Y., Liu, Q., Wang, G., Chen, E., & Zhang, H. (2017). P2P lending su ey: pla o ms, ecen ad ances and p ospec s. ACM T ansac ions on In elligen Sys ems and Technology (TIST), 8(6), 72. Zhao, D., Huang, C., Wei, Y., Yu, F., Wang, M., & Chen, H. (2017). An e ec i e compu a ional model o bank up cy p edic ion using ke nel ex eme lea ning machine app oach. Compu a ional Economics, 49(2), 325–341. h ps://doi.o g/10.1007/s10614-016-9562-7 48 8.3 APPENDIX C Table 11 - Desc ip i e S a is ics o Nume ical Va iables Sou ce: Au ho ’s p oduc ion. Va iables Mean S d.De Min Median Max A e age by loan s a us Fully Paid De aul Annual Income 73,250 49,039 3,000 62,000 1,000,000 74,563 65,787 Cha ged O wi hin 12 mon hs 0.01 0.11 0 0 10 0.01 0.01 C edi His o y 16.12 7.56 1 15 71 16.25 15.40 Delinquency pas 2 yea s 0.31 0.86 0 0 30 0.31 0.34 Deb - o-Income 17.64 8.20 0 17.15 41.21 17.36 19.20 FICO 695 31 640 690 845 697 687 Inqui ies las 6 mon hs 0.70 1 0 0 17 0.67 0.88 In e es a e 12.42 4.17 5.32 12.29 28.99 12.06 14.47 Loan Amoun 13,202 8,090 500 11,000 35,000 13,172 13,372 Open Accoun s 11.31 5.22 1 10 84 11.28 11.48 Public Reco ds 0.20 0.59 0 0 86 0.19 0.23 Public Reco d Bank up cies 0.12 0.36 0 0 12 0.12 0.14 Re ol ing Balance 16,032 22,095 0 11,053 2,904,836 16,289 14,573 Re ol ing line u iliza ion a e 54.07 23.95 0 54.80 892.30 53.56 56.98 Tax Liens 0.05 0.40 0 0 85.00 0.05 0.05 W i ing Skills 1.51 3.39 0 0 10 1.50 1.53 49 Table 12 - Desc ip i e S a is ics o Ca ego ical Va iables Va iables # issued loans (%) De aul a e (%) Employmen Leng h 0 – 4 yea s 277,107 (37.8%) 15.0% 5 – 9 yea s 182,924 (25.0% 15.0% 10+ yea s 231,027 (31.5%) 13.8% Missing 41,482 (5.7%) 21.0% G ade A 156,599 (21.4%) 5.5% B 238,806 (32.6%) 11.4% C 195,344 (26.7%) 18.5% D 94,012 (12.8%) 24.2% E 34,607 (4.7%) 29.8% F 10,974 (1.5%) 34.8% G 2,198 (0.3%) 37.9% Home Owne ship Mo gage 350,431 (47.8%) 12.8% Own 73,809 (10.1%) 15.5% Ren 308,300 (42.1%) 17.3% Pu pose Ca 8,371 (1.1%) 11.9% C edi ca d 173,984 (23.8%) 12.4% Deb consolida ion 418,877 (57.2%) 15.8% Educa ional 361 (0.05%) 20.2% Home imp o emen 42,686 (5.8%) 13.2% House 3,235 (0.4%) 18.9% Majo pu chase 15,314 (2.1%) 13.3% Medical 7,999 (1.1%) 17.5% Mo ing 5,195 (0.7%) 19.5% O he 39,582 (5.4%) 17.1% Renewable ene gy 554 (0.1%) 19.7% Small business 9,317 (1.3%) 24.4% Vaca ion 4,748 (0.6%) 16.1% Wedding 2,317 (0.3%) 12.3% Te m 36 mon hs 668,970 (91.3%) 14.0% 60 mon hs 63,570 (8.7%) 25.3% Ve i ica ion S a us Sou ce Ve i ied 255,845 (34.9%) 15.1% Ve i ied 242,057 (33.0%) 18.0% No Ve i ied 234,638 (32.0%) 11.7% Sou ce: Au ho ’s p oduc ion. Page | i