A machine learning approach for predicting bank credit worthiness

Abstract

Machine learning is an emerging technique for building analytic models for machines to "learn" from data and be able to do predictive analysis. The ability of machines to "learn" and do predictive analysis is very important in this era of big data and it has a wide range of application areas. For instance, banks and financial institutions are sometimes faced with the challenge of what risk factors to consider when advancing credit/loans to customers. For several features/attributes of the customers are normally taken into consideration, but most of these features have little predictive effect on the credit worthiness or otherwise of the customer. Furthermore, a robust and effective automated bank credit risk score that can aid in the prediction of customer credit worthiness very accurately is still a major challenge facing many banks. In this paper, we examine a real bank credit data and conduct several machine learning algorithms on the data for comparative analysis and to choose which algorithms are the best fit for learning bank credit data. The algorithms gave over 80% accuracy in prediction. Furthermore, the most important features that determine whether a customer will default or otherwise in paying his/her credit the next month are extracted from a total of 23 features. We then applied these most important features on some selected machine learning algorithms and compare their predictive accuracy with the other algorithms that used all the 23 features. The results show no significant di erence, signifying that these features can accurately determine the credit worthiness of the customers. Finally, we formulate a predictive model using the most important features to predict the credit worthiness of a given customer.