The goal of this research is supporting Borrower Risk Assessment in Peer-to-Peer (P2P) web-based Microfinance Platforms. Despite its recent fast growth in fame and money raised, P2P lending remains an understudied field in academia. P2P lending works through web platforms of financial transactions where borrowers place requests for loans and private lenders fund them directly or indirectly. These emerging platforms help microfinance overcome the challenge of sustainability with operational efficiency and cost effectiveness. What motivated us to focus on these P2P models is to support lenders in appraising the credit risk in choosing a borrower with the given available information. Historical data about loans, repayments, delays and delinquency situations are in fact made available by one of these platforms. This large amount of data is an asset that enables the development a support system using both expert knowledge and historical data through an AI approach. Available data describe loan episodes and their final outcomes, but they lack an indication of what should have been the suggested risk rating associated to the loan request. The main goal of this research is (a) build a Case Based Reasoning (CBR) system for borrower risk assessment in indirect P2P microfinance web platforms. Since the available loan episodes include a case description and outcome, but they lack a solution (risk rating) achieving the missing solution part in case represents another sub-objective: (b) to develop an expert-based risk rating model. This model has been used to effectively bootstrap the CBR system by producing a set of representative complete cases. We have chosen the CBR approach considering the unique nature of borrower profiles in microcredit, where specific characteristics demands for special knowledge. The proposed CBR system works as a statistical, incremental learning model to improve the results (risk rating) of expert rules, that were employed in the phase of initial cases definition. Expert knowledge is also instrumental in the definition of the similarity function guiding the retrieval of the CBR system. This study has brought to two contributions: expert-based risk rating and CBR-based risk rating. As a dominant risk rating approach in microfinance, expert-based risk rating approaches could assess borrower risk very well. However, they do not fit well online P2P situations due to inherent limitations like their static nature (no learning), which leads to the need of maintenance, often heavy computational costs, high user requirements and partial automation. Thus, the expert-based risk scoring has been used for providing the solution to a proper set of representative relevant cases to use in CBR-based risk rating which is instead completely automated, with an incremental learning, and in general more suited to an online P2P microcredit setting. We considered Kiva as the largest and leading P2P platform as a case study which allows to access its open source data for study. We recovered the Kiva XML data using XQuery to organize a database for past loans of individual borrowers with representative numbers and then we used examination method for identifying relevant and readily extractable features for the sample of past individual borrower loans. Using this expert-based models credit scoring has been carried out for a set of representative cases of 107 loans, and then they have been used in the CBR system as complete loan cases to run the system for assessing new loan applicants. The CBR rating has been tested with a set of test loan cases (75 cases from holdout sample from 2014) for evaluating its predictive power. The CBR system considered as low risk borrow requests 60 of them, 87% of which were correctly repaid; the system turned out to be quite conservative, since requests considered risky often turned out to be correctly repaid, but in general results are encouraging.

L’obiettivo della ricerca è supportare la valutazione del rischio in piattaforme web di microfinanza Peer-to-Peer (P2P). Nonostante la recente crescita di queste soluzioni, il contesto non è molto studiato nell’accademia. Il prestito tramite piattaforme web P2P prevede una transazione finanziaria tra un richiedente, che specifica un profilo e l’uso previsto per i fondi, e privati che finanziano prestiti direttamente o indirettamente. Queste piattaforme migliorano la sostenibilità della microfinanza con efficienza operativa a costi contenuti. La scelta di focalizzarci su questo contesto è motivata dall’idea di supportare i prestatori nella scelta dei prestiti da finanziare date le informazioni disponibili sui richiedenti. I dati storici su prestiti, rimborsi, ritardi e situazioni di default sono infatti resi disponibili da una delle piattaforme: Kiva. Questa ampia base di dati consente lo sviluppo di un sistema che usa sia conoscenza esperta che dati storici tramite un approccio di Intelligenza Artificiale. I dati descrivono episodi di prestito e l’esito finale, ma non un’indicazione del rischio che andrebbe associato alla richiesta. L’obiettivo principale della ricerca è quindi (a) costruire un sistema di Case Based Reasoning (CBR) per la valutazione del rischio associato a una richiesta di prestito in un sistema di microfinanza P2P. Dato che gli episodi includono una descrizione del caso ed il suo esito mancando però una soluzione, è stato affrontato un secondo problema: (b) sviluppare un sistema esperto di valutazione del rischio per produrre un insieme di casi rappresentativi nei quali la componente soluzione sia presente, generata tramite questo approccio. L’approccio CBR è motivato dalla natura dei profili di richiedenti prestito nel microcredito, dove specifiche caratteristiche richiedono una conoscenza particolare. Il sistema proposto combina un approccio statistico, con un apprendimento incrementale per migliorare i risultati delle regole degli esperti, che sono state usate per la definizione della base dei casi. La conoscenza esperta è anche stata fondamentale per la definizione della funzione di similarità che guida il recupero di casi simili nel sistema CBR. La ricerca ha prodotto due contributi: una tecnica di valutazione del rischio basata su conoscenza esperta ed una basata su un approccio CBR. La prima tecnica ha fornito finora risultati molto positivi, ma non è adatta allo scenario dei prestiti online P2P date le inerenti limitazioni come la natura statica dell’approccio, che richiede manutenzione e aggiornamento manuale, costi computazionali a volte elevati, requisiti utente elevati e non sempre completa automazione. Questo approccio è stato quindi utilizzato per generare un insieme di casi rappresentativi completi da usare per abilitare il sistema CBR, che è invece completamente automatizzato, dotato di una forma di apprendimento incrementale, e in genere più adatto al contesto del microcredito online. Abbiamo considerato Kiva in quanto si tratta della piattaforma P2P più grande e in quanto consente un accesso completo alle informazioni necessarie per questo studio. I dati XML sono stati analizzati tramite XQuery per organizzare un database di prestiti passati e richiedenti prestito, con numeri sufficientemente rappresentativi, e un’analisi è stata effettuata per identificare le caratteristiche più rilevante dei richiedenti prestito. Usando il sistema esperto 107 casi rappresentativi sono stati definiti per la base dei casi. Il sistema CBR è stato validato usando 75 ulteriori episodi dei quali era noto l’esito finale. Per 60 di questi episodi il sistema ha dato una valutazione di basso rischio e nell’87% dei casi la scelta è stata corretta (prestiti ripagati correttamente); il sistema è invece stato eccessivamente cauto nel caso di valutazioni di alto rischio, ma i risultati sono complessivamente incoraggianti.

(2017). Borrower Risk Assessment in P2P Microfinance Platforms. (Tesi di dottorato, Università degli Studi di Milano-Bicocca, 2017).

Borrower Risk Assessment in P2P Microfinance Platforms

JAMAL UDDIN, MOHAMMED
2017

Abstract

The goal of this research is supporting Borrower Risk Assessment in Peer-to-Peer (P2P) web-based Microfinance Platforms. Despite its recent fast growth in fame and money raised, P2P lending remains an understudied field in academia. P2P lending works through web platforms of financial transactions where borrowers place requests for loans and private lenders fund them directly or indirectly. These emerging platforms help microfinance overcome the challenge of sustainability with operational efficiency and cost effectiveness. What motivated us to focus on these P2P models is to support lenders in appraising the credit risk in choosing a borrower with the given available information. Historical data about loans, repayments, delays and delinquency situations are in fact made available by one of these platforms. This large amount of data is an asset that enables the development a support system using both expert knowledge and historical data through an AI approach. Available data describe loan episodes and their final outcomes, but they lack an indication of what should have been the suggested risk rating associated to the loan request. The main goal of this research is (a) build a Case Based Reasoning (CBR) system for borrower risk assessment in indirect P2P microfinance web platforms. Since the available loan episodes include a case description and outcome, but they lack a solution (risk rating) achieving the missing solution part in case represents another sub-objective: (b) to develop an expert-based risk rating model. This model has been used to effectively bootstrap the CBR system by producing a set of representative complete cases. We have chosen the CBR approach considering the unique nature of borrower profiles in microcredit, where specific characteristics demands for special knowledge. The proposed CBR system works as a statistical, incremental learning model to improve the results (risk rating) of expert rules, that were employed in the phase of initial cases definition. Expert knowledge is also instrumental in the definition of the similarity function guiding the retrieval of the CBR system. This study has brought to two contributions: expert-based risk rating and CBR-based risk rating. As a dominant risk rating approach in microfinance, expert-based risk rating approaches could assess borrower risk very well. However, they do not fit well online P2P situations due to inherent limitations like their static nature (no learning), which leads to the need of maintenance, often heavy computational costs, high user requirements and partial automation. Thus, the expert-based risk scoring has been used for providing the solution to a proper set of representative relevant cases to use in CBR-based risk rating which is instead completely automated, with an incremental learning, and in general more suited to an online P2P microcredit setting. We considered Kiva as the largest and leading P2P platform as a case study which allows to access its open source data for study. We recovered the Kiva XML data using XQuery to organize a database for past loans of individual borrowers with representative numbers and then we used examination method for identifying relevant and readily extractable features for the sample of past individual borrower loans. Using this expert-based models credit scoring has been carried out for a set of representative cases of 107 loans, and then they have been used in the CBR system as complete loan cases to run the system for assessing new loan applicants. The CBR rating has been tested with a set of test loan cases (75 cases from holdout sample from 2014) for evaluating its predictive power. The CBR system considered as low risk borrow requests 60 of them, 87% of which were correctly repaid; the system turned out to be quite conservative, since requests considered risky often turned out to be correctly repaid, but in general results are encouraging.
VIZZARI, GIUSEPPE
BANDINI, STEFANIA
Microfinance,; P2P_Lending,; CBR,; Expert_Rules,; Credit_Scoring
Microfinance,; P2P_Lending,; CBR,; Expert_Rules,; Credit_Scoring
INF/01 - INFORMATICA
English
27-mar-2017
INFORMATICA - 87R
29
2015/2016
open
(2017). Borrower Risk Assessment in P2P Microfinance Platforms. (Tesi di dottorato, Università degli Studi di Milano-Bicocca, 2017).
File in questo prodotto:
File Dimensione Formato  
phd_unimib_787876.pdf

accesso aperto

Descrizione: tesi di dottorato
Tipologia di allegato: Doctoral thesis
Dimensione 3.34 MB
Formato Adobe PDF
3.34 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/10281/151637
Citazioni
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
Social impact