Predicting loan default in peer‐to‐peer lending using narrative data

• Published date: 01 March 2020
• Date: 01 March 2020
• Author: Lingyun He,Yanlin Ding,Yufei Xia,Yinguo Li,Nana Liu
• DOI: http://doi.org/10.1002/for.2625

RESEARCH ARTICLE

Predicting loan default in peer‐to‐peer lending using

narrative data

Yufei Xia

1

| Lingyun He

2

| Yinguo Li

1

| Nana Liu

2

| Yanlin Ding

3

1

Business School, Jiangsu Normal

University, Xuzhou, Jiangsu, China

2

School of Management, China University

of Mining and Technology, Xuzhou,

Jiangsu, China

3

Faculty of Economics and Management,

Shangluo University, Shangluo, Shaanxi,

China

Correspondence

Yufei Xia, Business School, Jiangsu

Normal University. Xuzhou, Jiangsu

221116, China.

Email: 6020180093@jsnu.edu.cn

Funding information

National Natural Science Foundation of

China, Grant/Award Numbers: 71874185

and 71573110; National Social Science

Fund Project of China, Grant/Award

Number: 15BTJ033; Ministry of Education

of Humanities and Social Science Project

of China, Grant/Award Number:

16YJAZH015; Research Support Project

for Doctoral Degree Teachers of Jiangsu

Normal University, Grant/Award Num-

ber: 18XWRX021

Abstract

Peer‐to‐peer (P2P) lending is facing severe information asymmetry problems

and depends highly on the internal credit scoring system. This paper provides

a novel credit scoring model, which forecasts the probability of default for each

applicant and guides the lenders' decision‐making in P2P lending. The pro-

posal is expected to improve the existing credit scoring models in P2P lending

from two aspects, namely the classifier and the usage of narrative data. We uti-

lize an advanced gradient boosting decision tree technique (i.e., CatBoost) to

predict default loans. Moreover, a soft information extraction technique based

on keyword clustering is developed to compensate for the insufficient hard

credit data. Validated on three real‐world datasets, the experimental results

demonstrate that variables extracted from narrative data are powerful features,

and the utilization of narrative data significantly improves the predictability

relative to solely using hard information. The results of sensitivity analysis

reveal that CatBoost outperforms the industry benchmark under different clus-

ter numbers of extracted soft information; meanwhile a small number of clus-

ters (e.g., three) is preferred for consideration of model performance,

computational cost, and comprehensibility. We finally facilitate a discussion

on practical implication and explanatory considerations.

KEYWORDS

credit scoring, narrative data, P2P lending, probability of default, soft information

1|INTRODUCTION

Peer‐to‐peer (P2P) lending, a type of crowdfunding in a

border sense, is emerging quickly across the world (Short,

Ketchen, McKenny, Allison, & Ireland, 2017). Bypassing

commercial banks, the borrowers and lenders are

matched via online platforms. Concretely, the borrowers

submit a loan application on the platforms and then the

potential lenders can freely browse the websites and

finally decide which one to invest in (bidding), as well

as the bidding amount. Owing to the absence of commer-

cial banks, P2P lending leads to lower overhead costs and

attracts individuals and small business (Guo, Zhou, Luo,

Liu, & Xiong, 2016).

Similar to traditional consumer lending; however, P2P

lending is facing severe information asymmetry problems

(Lee & Lee, 2012). The possible reasons are as follows:

First, in most P2P lending platforms, borrowers and

lenders are usually anonymous and matched via online

platforms. Although part of credit information may be

available to the public, the anonymity may cause infor-

mation failure in P2P lending (Freedman & Jin, 2017).

Second, the unskilled and unprofessional lenders in P2P

lending may fail to screen the creditworthiness of

Received: 2 January 2019 Revised: 3 June 2019 Accepted: 24 July 2019

DOI: 10.1002/for.2625

Journal of Forecasting. 2020;39:260–280.wileyonlinelibrary.com/journal/for© 2019 John Wiley & Sons, Ltd.260

borrowers properly by employing quantifiable informa-

tion (Mild, Waitz, & Wöckl, 2015). Finally, the P2P lend-

ing is usually not secured by collateral. Consequently, the

lenders would have little knowledge on the ability and

willingness to pay for a specific loan application.

To diminish the risk of information failure, the plat-

forms have employed several measures at present and

probably the most popular one is the internal credit

scoring system. From the perspective of method, the

credit scoring model inherently estimates probability of

default (PD) to predict loan default. In most platforms,

each borrower is assigned a credit grade based on the

hard credit data of the borrowers. These data mainly

include information concerning solvency, demographic

variables, and loan characteristics. The internal credit

scoring is regarded as an efficient decision support sys-

tem for lenders, and many studies have confirmed the

effects of internal credit grade on lenders' behavior and

predicting defaults (Iyer, Khwaja, Luttmer, & Shue,

2015; Malekipirbazari & Aksakalli, 2015; Xia, Liu, &

Liu, 2017).

Enhancing the efficiency of predicting default loans is

the permanent pursuit of credit risk management. Even a

slight improvement in predict potential default may lead

to huge economic benefits (G. Wang, Hao, Ma, & Jiang,

2011). To improve the predictive performance of the

credit scoring model, current literature mainly contrib-

utes from two aspects, namely method and data. Regard-

ing the classification algorithms, logistic regression (LR)

(Emekter, Tu, Jirasakuldech, & Lu, 2015), neural network

(NN) (Byanjankar, Heikkilä, & Mezei, 2015), random for-

ests (RF) (Malekipirbazari & Aksakalli, 2015; Z. Wang,

Jiang, Ding, Lv, & Liu, 2018), gradient boosting decision

tree (GBDT) (Liu & Wu, 2017), and deep learning

approach (Luo, Wu, & Wu, 2017) have been applied to

predict the default loans. Moreover, Serrano‐Cinca and

Gutiérrez‐Nieto (2016), Guo et al. (2016), and Xia, Liu,

and Liu (2017) have proposed profit scoring models to

guide lenders' investment through portfolio optimization

strategy.

From the perspective of credit‐related data, most stud-

ies calculated credit scores based on hard information.

Some scholars highlighted the sample bias caused by

the large amount of reject application in P2P lending

and proposed reject inference techniques to infer the true

loan status of rejected applications and integrate them

into credit risk assessment (Li, Tian, Li, Zhou, & Yang,

2017; Xia, Yang, & Zhang, 2018). Some interesting studies

utilized soft information to boost the predictability of

credit scoring models. Though many studies have

discussed the usage of soft and hard information,

there is no precise definition of soft information at pres-

ent. Liberti and Petersen (2019) summarized the

characteristics of soft and hard information to make them

classifiable. These characteristics include the following:

1. Soft information is often communicated as text,

whereas hard information is typically recorded in

numbers.

2. The context in soft information is more important rel-

ative to that of hard information.

3. The collector and decision maker of soft information

is often the same person, whereas the hard informa-

tion collection and decision making are separated;

4. Soft information is typically gauged in terms of verifi-

ability, whereas some hard information is verifiable.

This is one of the defining characteristics of soft

information in P2P lending, since some information

self‐reported by borrowers is usually unverifiable

(Michels, 2012).

Some scholars have attempted to use soft information in

the credit risk assessment of P2P lending. Social network

and descriptive text (narratives) remain two major

sources of soft information. Concerning social network,

Lin, Prabhala, and Viswanathan (2013) revealed that

friendships positively affected the probability of loan

funding and decreased PD. Moreover, Wei, Yildirim,

Van den Bulte, and Dellarocas (2015) developed a variety

of methods to examine the effect of social network data in

predicting potential default borrowers. The results

revealed that social network data can improve the perfor-

mances of credit scores in some cases. X. Chen, Zhou,

and Wan (2016) observed that the borrower's social capi-

tal had a considerable impact on repayment. Regarding

narratives, Dorfleitner et al. (2016) found text description

affected funding probability but hardly predict PD in P2P

lending. Utilizing data from a Chinese representative

platform, X. Chen, Huang, and Ye (2018) claimed that

the amount of punctuation in loan descriptions affected

both funding probability and interest rate. Jiang, Wang,

Wang, and Ding (2018) proposed a credit scoring model

integrating soft information (extracted from narratives)

and hard credit data. The empirical analysis showed that

the proposal boosted the predictive performance relative

to methods solely using hard credit data. Furthermore,

Yao, Chen, Wei, Chen, and Yang (2018) examined the

role of soft information contained in loan titles for online

lending.

Although abundant studies have focused on credit

scoring in P2P lending, some improvements can be made

from the following aspects. First, given that ensemble

techniques such as RF and GBDT have been applied in

credit scoring of P2P lending, powerful ensemble classi-

fiers are always welcome in this area. Second, some appli-

cants seldom have social media account, which hinders

XIA ET AL.261