Personal Credit Risk Management System for Sri Lankan Banks

Abstract

The financial sector and banks are the leaders of a country's economic development. Banks set immense effort into managing credit losses as it causes a significant impact on a bank's performance. The latest business patterns associated with digitalization have altered the prevailing credit risk functions. Increased demand for automation and digitalization in banks create exabytes of structured and unstructured data, imposing an urgent necessity of upgrading credit risk management systems. Many countries successfully utilize big data and deep learning techniques for default predictions. Hence, countries like Sri Lanka have the opportunity to upgrade default prediction systems with modern techniques. Deep learning techniques have drawn attention in banking credit risk analysis compared to traditional machine learning techniques due to its training and learning capacity. This study aimed to develop a personal credit risk prediction system using a modified RNN-LSTM model. First, the irrelevant columns and data in the dataset were removed. RandomforestClassifier was used to avoid overfitting and to measure feature importance. After completion of the data set cleaning process, implementation of the proposed model was initiated. The proposed LSTM model has two LSTM layers with a dropout of 0.2 rates at the first LSTM layer. Dense layer with 25 nodes, Dense layer with 1 node, and sigmoid function. The model had an RMSE of 0.44 and a prediction accuracy of 73.61%. Hence, it is substantiated that there is a high possibility to employ LSTM architecture for the defaulter prediction system.