A Comparative Study between Machine and Deep Learning Models for the Prediction of Bank Credit Recovery

Abstract

Nowadays, technology is advancing rapidly. With the advancement of technology, many institutions are adapting their business with new technologies. Institutions have huge amounts of data about the employers and clients. To handle huge amounts of data, many institutions are applying many data mining techniques to maintain their institutions properly and fast. In financial institutions like banks, to analyze and handle the data about the customer is very necessary. To analyze the credit risk is a primary field in the banking sectors and there are many techniques exist to predict whether a customer is credit worthy or not and the possibility of loan default. In this research, we’ve used a dataset from a Bangladeshi bank. The dataset is the credit defaulter dataset. We tried to predict the delinquent customers who have the highest possibility of short term credit recovery. We applied some machine and deep learning models to predict the credit recovery. The dataset is imbalanced. First of all we balanced the dataset by using SMOTE technique and then we performed feature scaling, feature selection process on the dataset. Finally, we applied machine and deep learning models. Compared with all of the models, Random Forest (RF) performed better than other models. We applied those models in both Train Test Split and Stratified K-Fold CV methods. In the Train Test Split method, RF gives 93% accuracy and in the Stratified K-Fold CV method, RF gives 94% accuracy. The result of the evaluation and statistical metrics of this model are also good in both of these methods. In the case of deep learning models, the best output comes from Artificial Neural Network (ANN) and Multilayer Perceptron (MLP) with 90% accuracy. Overall RF performed better and can better predict the credit recovery.