Predicting credit card delinquencies: An application of deep neural networks

• Date: 01 October 2018
• DOI: http://doi.org/10.1002/isaf.1437
• Published date: 01 October 2018
• Author: Ting Sun,Miklos A. Vasarhelyi

RESEARCH ARTICLE

Predicting credit card delinquencies: An application of deep

neural networks

Ting Sun

1

|Miklos A. Vasarhelyi

2

1

The College of New Jersey, Accounting and

Information Systems, Ewing, NJ, USA

2

Rutgers University, Newark Business School,

Accounting and Information Systems, Newark,

NJ, USA

Correspondence

Ting Sun, The College of New Jersey,

Accounting and Information Systems, Ewing,

NJ, USA.

Email: tsun9920@gmail.com

Summary

The objective of this paper is twofold. First, it develops a prediction system to help

the credit card issuer model the credit card delinquency risk. Second, it seeks to

explore the potential of deep learning (also called a deep neural network), an emerging

artificial intelligence technology, in the credit risk domain. With real‐life credit card

data linked to 711,397 credit card holders from a large bank in Brazil, this study

develops a deep neural network to evaluate the risk of credit card delinquency based

on the client's personal characteristics and the spending behaviours. Compared with

machine‐learning algorithms of logistic regression, naive Bayes, traditional artificial

neural networks, and decision trees, deep neural networks have a better overall

predictive performance with the highest Fscores and area under the receiver

operating characteristic curve. The successful application of deep learning implies that

artificial intelligence has great potential to support and automate credit risk assess-

ment for financial institutions and credit bureaus.

KEYWORDS

artificial intelligence, credit carddelinquency, deep neural network, machine learning, risk

assessment

1|INTRODUCTION

Credit card debt is climbing rapidly. For example, the total amount of

outstanding revolving credit debt in the USA reached more than $1

trillion in 2017, according to data from the US Federal Reserve

(https://www.federalreserve.gov/releases/g19/hist/cc_hist_sa_levels.

html). This is the highest level of credit card debt since January 2009

(Porche, 2017). In the UK, the annual rate of credit card lending in

September 2017 expanded by 9.2% on the same month a year earlier

(Chu, 2017). This trend serves as an alarm for a high risk of credit card

delinquencies. The S&P/Experian Bankcard Default Index shows that

the credit card delinquency rate in the USA in March 2017 had reached

its highest point since June 2013 (3.31%; Durden, 2017). Currently, the

credit risk assessment has become a critical basis for the credit card

issuer's decision‐making. The failure of effectively evaluating the risk

could result in a high nonperforming ratio, increased debt collection

cost, and growing bad debt counts, which threaten the health of the

credit card industry (Chen & Huang, 2011; Twala, 2010).

Owing to the significance of credit risk (especiallythe delinquency

risk) assessment, many techniqueshave been proposed with promising

results; for example, Discriminantanalysis and logistic regression, deci-

sion trees, and support vector machine (SVM; Marqués, García, &

Sánchez, 2012). Artificial neural networks (ANNs)were also employed

to forecast credit risk (Koh & Chan, 2002;Thomas, 2000). Over the past

decade, deep learning (also called a deep neural network (DNN)), an

emerging artificial intelligence technique, has been applied to a variety

of areas.It achievedexcellent predictive performance in areas like health

care and computer games, where data are complex and big (Hamet &

Tremblay,2017) and exhibited great potential to be used for many other

areas where human decision‐making is inadequate (Ohlsson, 2017).

However, this approach has not been applied to predict credit card

delinquencies.Furthermore, it is unclear whether this approach is supe-

rior to other machine‐learning approaches. Today's business becomes

more and more complex. The scale and complexity of the data make

the decision‐makingof financial institutions much more challengingthan

ever before, even with the help of traditionaldata analytical technology

Received: 15 February 2018 Revised: 26 June 2018 Accepted: 26 June 2018

DOI: 10.1002/isaf.1437

174 © 2018 John Wiley & Sons, Ltd. Intell Sys Acc Fin Mgmt. 2018;25:174–189.wileyonlinelibrary.com/journal/isaf

(Turban, Aronson, Liang, & McCarthy, 2005). Hence, it is necessary to

apply this state‐of‐the‐arttechnology to develop intelligent systems to

support and automate the decision‐makingof credit card issuers based

on large volumes of data (Ohlsson, 2017).

This paper bridges this gap by developing a DNN for the predic-

tion of credit card delinquency risk. Using real‐life credit card data

from a large bank in Brazil, this study demonstrates the effectiveness

of DNN in assisting financial institutions to quantify and manage the

credit risk for the decision‐making of credit card issuance and loan

approval. Prior research suggests that financial statements and

customers' transactional records are useful for credit risk assessment

(Yeh & Lien, 2009). As for credit card delinquency, researchers

associated it with the personal characteristics of credit card holders

and their spending behaviours (e.g. Chen & Huang, 2011; Khandani,

Kim, & Lo, 2010). To date, as data storage becomes more convenient

and inexpensive, financial institutions have accumulated a massive

amount of data about credit card transactions and clients' personal

characteristics. This provides an excellent opportunity to use deep

learning to establish delinquency risk prediction models. This paper

constructs models using the personal information of the credit card

holder (e.g. occupation, age, and region of residence), the accumulative

transactional information (e.g. the frequency that the client has been

billed, and the total amount of cash withdrawals) based on the bank's

record in September 2013, and the data of transactions that occurred

in June 2013. After comparing deep learning with other machine‐

learning algorithms with regard to the predictive performance on our

data, we find that the DNN outperforms other models in terms of

better F

1

and area under the receiver operating characteristic (ROC)

curve (AUC), which measure the overall predictive accuracy. The result

suggests that deep learning can be a useful addition to the current

toolkit for credit risk assessment, especially for financial institutions

and regulators in the credit card industry, and especially for data with

severe imbalance issues, large size, and complex structure.

The remainder of the paper is organized as follows. Section 2

reviews the prior literature and addresses the research gap. Section 3

overviews the deep learning method and discusses the differences

between deep learning and other machine‐learning approaches.

Section 4 introduces the data and variables. The modelling process

and the results are presented in Sections 5 and 6 respectively.

Section 7 concludes the paper and discusses the limitations and

future research.

2|LITERATURE REVIEW

Prior research has proposed a variety of data‐mining techniques to pre-

dict credit card delinquencies. Those techniques include statistical

modelling approaches, such as discriminant analysis, logistic regression

and k‐nearest neighbours (KNN; Abdou & Pointon, 2011), and machine‐

learning approaches, such as decision trees, naive Bayes, and SVMs. In

addition, ANNs are considered an important alternative method.

2.1 |Statistical Modelling

The standard approach of estimating the probability of credit card

delinquencies is logistic regression (Crook, Edelman, & Thomas,

2007; Kruppa, Schwarz, Arminger, & Ziegler, 2013). Wiginton (1980)

proposed one of the earliest studies comparing logistic regression with

discriminant analysis for credit scoring. It found that the logistic

regression model exhibited a better accuracy rate than the model of

discriminant analysis. Leonard (1993) used logistic regression with ran-

dom effects to evaluate commercial loans for a major Canadian bank.

Logistic regression was also applied by Abdou, Pointon, and El‐Masry

(2008), who investigated the credit risk for Egyptian banks with logis-

tic regression, discriminant analysis, probit analysis, and ANNs. How-

ever, it was argued that the underlying assumptions of logistic

regression are rather strict (Malley, Kruppa, Dasgupta, Malley, &

Ziegler, 2012). For instance, multicollinearity should not exist among

independent variables.

2.2 |Machine Learning

Unlike statistical modelling, which has predefined structures and

assumptions, machine learning allows the computer to learn the partic-

ular structure of the model from the data (Huang, Chen, Hsu, Chen, &

Wu, 2004). As a result, another research stream explored the problem

of credit risk assessment by using machine‐learning algorithms. With

three datasets from UCI Irvine Machine Learning Database, Lahmiri

(2016) constructed five machine‐learning models, including an SVM, a

backpropagation neural network, a radial basis function neural net-

work, linear discriminant analysis, and naive Bayes, to assess credit risk.

They found that the SVM provides the best predictive accuracy for all

three datasets. Using credit bureau, transactional, and account‐balance

data from January 2005 to April 2009 of a major commercial bank,

Khandani et al. (2010) employed generalized classification and regres-

sion trees that was initially proposed by Breiman, Friedman, Olshen,

and Stone (1984) to forecast the consumers' delinquencies. The

authors asserted that the classification and regression trees method

was superior to logistic regression, discriminant analysis models, and

credit scores with regard to identifying subtle nonlinear relationships

underlying the massive dataset. Butaru et al. (2016) analysed another

large sample of account‐level credit card data from six major commer-

cial banks from January 2009 to December 2013. They developed and

compared prediction models for credit card delinquencies with logistic

regression, a decision tree (C4.5), and a random forest. It was concluded

that, although all models performed reasonably well, the decision tree

and the random forest outperformed the logistic regression.

Other studies investigated credit default. For example, Kruppa

et al. (2013) applied random forest, optimized logistic regression,

KNN, and bagged KNN to estimate the credit default probability to a

large dataset of short‐termed instalment credits for a company pro-

ducing household appliances. This study demonstrated the superiority

of random forest over other models. The data set consisted of 64,524

transactions, with 13% of the total financed amounts remain

uncollectable (which means default). However, in a real‐world case,

the ratio of credit card default is usually much lower than 13%.

Fitzpatrick and Mues (2016) focused on the mortgage default. They

applied boosted regression trees, random forests, penalized linear

and semi‐parametric logistic regression models to four portfolios of

over 300,000 Irish owner‐occupier mortgages. The results showed

that although those models had varying degrees of predictive

SUN AND VASARHELYI 175