Classification of banking clients according to their loan default status using machine learning algorithms.

Abstract

Loan lending has become crucial for both individuals and companies. For lending institutions, although profitable, it can be very risky due to clients defaulting on their loan agreement. Credit risk assessment is a critical process which is carried out by most lending institutions; it reduces the possibility of lending to clients who will default on their loan repayment, however, it does not eliminate the problem. Thus, a collections process which aims to retrieve unpaid debt is also necessary. With South Africa facing another recession, which was only worsened by the lockdown during the covid-19 pandemic, lending institutions can expect an increase in the number loan defaulters. To counter this increase, changes will have to be made to their policies and processes. Changes can be made to either the loan application procedures (e.g. credit risk assessment, affordability assessment et cetera) or the post disbursal procedures (e.g. collections processes). The aim of this study is to predict whether a client will default on his/her loan, using machine learning algorithms, in order to enhance the collection process of the financial institution under study, where default is defined as missing at least three payments in the first 12 months of the loan being granted. The logistic regression model, decision tree, random forest, support vector machine, Naïve Bayes classifier, k-nearest neighbours algorithm and the artificial neural network were fitted to the balanced dataset. In the researcher’s analysis, loan data from a South African financial institution were used for the period August 2019 to December 2019. Variables related to a client’s demographics, income, expenses and debt, as well as loan information, were included in the dataset. Exploratory data analysis (EDA) was utilised in order to analyse the dataset and summarise their main characteristics. To reduce the dimensionality of the dataset, two techniques were used, namely principal component analysis (PCA), which is also used to correct the data for multicollinearity, and feature selection (i.e., recursive feature elimination). Each model was fitted to the dataset using these two techniques, and the confusion matrix and metrics such balanced accuracy, true positive ratio, true negative ratio, AUC score and the Gini coefficient were used to evaluate the different models in order to determine which model performed the best and was most suited for this application problem. The results show that when using the PCA approach, the random forest model, which obtained a balanced accuracy score, true positive ratio and AUC score of 0.69, 0.74 and 0.74, respectively, performed the best. The random forest model also performed the best when using the feature selection technique, obtaining a balanced accuracy score, true positive ratio and AUC score of 0.69, 0.74 and 0.75, respectively. When comparing the random forest model using PCA to the random forest model using feature selection, the results showed a marginal difference between each performance metric analysed. The random forest model using PCA utilised 48 variables, whereas the random forest model using feature selection utilised only 18 variables and thus seemed to be more suitable for the classification problem under study. The results of this study are expected to benefit analysts and data scientists in financial institutions who would like to identify the robust machine learning algorithms for classifying defaulting clients. This study is also of significance to policy makers who would want to identify the risk factors associated with loan defaulting clients.