Jozef Zurada

The failure or success of the banking industry depends largely on the industry’s ability to properly evaluate credit risk. In the consumer-lending context, the bank’s goal is to maximize income by issuing as many good loans to consumers as possible while avoiding losses associated with bad loans. Mistakes could severely affect profits because the losses associated with one bad loan may undermine the income earned on many good loans. Therefore banks carefully evaluate the financial status of each customer as well as their credit worthiness and weigh them against the banks’ internal loan-granting policies. Recognizing that even a small improvement in credit scoring accuracy translates into significant future savings, the banking industry and the scientific community have been employing various machine learning and traditional statistical techniques to improve credit risk prediction accuracy. This paper examines historical data from consumer loans issued by a financial institution to individuals that the financial institution deemed to be qualified customers. The data consists of the financial attributes of each customer and includes a mixture of loans that the customers paid off and defaulted upon. The paper uses three different data mining techniques (decision trees, neural networks, logit regression) and the ensemble model, which combines the three techniques, to predict whether a particular customer defaulted or paid off his/her loan. The paper then compares the effectiveness of each technique and analyzes the risk of default inherent in each loan and group of loans. The data mining classification techniques and analysis can enable banks to more precisely classify consumers into various credit risk groups. Knowing what risk group a consumer falls into would allow a bank to fine tune its lending policies by recognizing high risk groups of consumers to whom loans should not be issued, and identifying safer loans that should be issued, on terms commensurate with the risk of default.