Research on the Factors that Influence P2P Loan Default: Analysis Based on Logistic Regression and Machine Learning
|Subject category of dissertation|
|Place of Publication|
In recent years, P2P loans have developed rapidly in developing countries, making great contributions to reducing the poverty rate of the local population, supporting the development of SMEs, and improving the level of financial inclusion. However, there are also many problems in the development of P2P loans, for example, information asymmetry, and the default rate of P2P loans is significantly higher than that of bank loans, which may disrupt the financial market order. To relieve the information asymmetry in P2P loans, this paper proposes that the borrower's mobile phone usage patterns and bank transaction patterns, which are accessible to the P2P lending platforms predict the borrower's default probability. This paper uses a set of operational data from a P2P lending platform in Indonesia and estimates the impact of mobile phone usage patterns and bank transaction patterns on the borrower's default probability by constructing a Logistic Regression model. Moreover, this paper has conducted predictions on the default probability of a transaction through three machine learning models, including the Random Forest model, XGBoost model, and Deep Neural Network model respectively.
This paper has found that the mobile phone usage patterns and the bank transaction patterns of the borrowers could significantly influence the default rate of their P2P lending loans. From the perspective of phone calls, borrowers who have received phone calls from more people, shorter average call time and fewer night calls are less likely to default. In terms of phone recharge patterns, borrowers who recharge their mobile phone more frequently with a small amount have a lower default risk. In the view of social media, borrowers who have installed WhatsApp are less likely to default. For the bank transaction patterns, borrowers who have more bank transaction records and higher average transaction amounts have lower default risk. Moreover, with the factors mentioned above, high-precision prediction of the default risk of a loan transaction could be achieved through the machine learning models, among which the XGBoost model is superior to the other two models in both prediction accuracy and model stability. Therefore, the XGBoost is the preferred model.
This paper has a significant theoretical contribution and practical significance. From the perspective of theoretical contribution, this paper introduces some new predictors for the analysis of the factors that would influence the P2P loan default rate, so that the relevant research will no longer be limited to the personal information provided by the users, instead, factors that the P2P lending platforms could collect actively could be applied.
From a practical point of view, the research results of this paper point out that P2P lending platforms in developing countries can foresee the default risk of a transaction by using the relevant data of the borrower's mobile phone usage patterns and bank transaction patterns, prevent the occurrence of ultra-high-risk transactions, reduce the platform risk, ensure the safety of investors' funds so that the P2P platforms can develop sustainably and contribute to the realization of inclusive finance.
|Year of Degree Awarded|
 ZHOU W, ARNER D W, BUCKLEY R P. Regulating FinTech in China: From Permissive to Balanced [J]. 2017.
|Academic Degree Assessment Sub committee|
|Domestic book classification number|
|Department||Department of Information Systems and Management Engineering|
Peng XR. Research on the Factors that Influence P2P Loan Default: Analysis Based on Logistic Regression and Machine Learning[D]. 深圳. 南方科技大学,2022.
|Files in This Item:|
|12032752-彭筱茹-商学院.pdf（3885KB）||Restricted Access||--||Fulltext Requests|
|Recommend this item|
|Export to Endnote|
|Export to Excel|
|Export to Csv|
|Similar articles in Google Scholar|
|Similar articles in Baidu Scholar|
|Similar articles in Bing Scholar|
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.