Bank Loan Approval Prediction Using Machine Learning Techniques

Abstract

Loan is the essential product of banks and other financial institutions. As a big number of people go to banks to borrow money for different activities, the number of customers have increased and some banks expect to earn a lot of money as a result of interest paid on loans. However, loans are associated with risk of defaulting, i.e. the possibility that some borrowers may not be able to pay back their loans. Thus, high levels of non-performing loans can be a source of instability of the banking sector and lead to bankruptcy. One of the important steps for banks to decide if a loan has to be authorized is to ensure that the candidate to borrow has the capacity of paying back the loan in the proposed terms. The advancement of technology like machine learning, computer science and other science is playing an important role by supporting banks to predict the probability of defaulting for a given customer based on his past behavior. In this research, we contribute to work by commercial banks to predict the behaviors of borrowers by developing and testing the accuracy of different models using data from Bank of Kigali. Collected data was divided into training dataset and test dataset where the train dataset was made by 70% and 30% was for test. After training the machine by using the training dataset, then we used the test dataset to check the accuracy of different models. By running ensembles, combinations of different machine learning techniques are used to find the best to use while predicting bank loans default prediction. The results of our analysis show that the Gradient Boosting is the best model to predict bank loan default, followed by XGBoosting while others like decision trees, random forest, logistic regression performed poorly. I would recommend financial institutions to use machine learning techniques because it saves money and time for both sides. Moreover, Findings show that the customers with Credit Score B will have low probability of defaulting. In this work we used data from one financial institution and we would recommend anyone who might want to further this study to consider using data from different financials institution across the region to capture the insight.