A Machine learning approach to assist the prediction of loan characteristics

Abstract

The business environment in Sri Lanka has become complex and competitive with the development of the financial sector and the spread of the Covid-19 pandemic. The number of business organizations and individuals applying for loans has increased. The practices that are being used to predict financial allocation for loans of future periods are based on previous experiences and rough estimates. The most challenging risk faced during this process is the credit risk, which is the risk of lending money to unsuitable loan applicants. Lengthy authentication procedures are being followed by financial institutes prior to approving loans. However, there is no assurance whether the chosen applicant is the right applicant or not. Also, predicting the risks of credit loans prior to becoming non-performing is essential as the outcomes are unbearable except provisions are arranged for anticipated downsides. Thus, this study focused on analyzing the historical data of loans and evaluating customer profiles based on the demographic, geographical, and behavioral data of the customers to enable the prediction of future loan amounts, evaluation of the credit risks of loans and prediction of Non-Performing Loans using Machine Learning (ML) algorithms, in order to help make appropriate choices in the future. An exploratory data analysis was first performed to provide insights on developing marketing strategies based on loan types and to identify the type of customers who can be approached. Thus, three models were devised to predict the identified loan characteristics. Model 1 was devised to predict the future loan amounts with the highest R-squared score of 0.9967 using Light Gradient Boosting Regression. Model 2 was devised to evaluate the credit risk with the highest training and test accuracy of 0.9960 and 0.7842, respectively, using Stacking Ensemble Classification. Model 3 was devised to predict the Non-Performing Loans with the highest training and test accuracy of 0.9999 and 0.9522, respectively, using Random Forest Classification. Finally, the study illustrated a remarkable approach in predicting loan characteristics which ideally suits the ever changing economy. It achieved outstanding results which could enable any financial institute in the country, in minimizing the expected risks.