Fraud Risk Prediction in Merchant-Bank Relationship using Regression Modeling

Banking industry has gone through one of the worst crisis in recent times, and is still recovering from the after-shocks. However, there were a lot of learnings that banks would have taken away from this crisis. One of them is the need for a robust risk management system. The crisis dealt a blow to the banking system, catching them off guard when it came to foreseeing the risk. Banks, in the credit card business, face financial risk in the form of both credit risk and fraud risk. Sharma and Agarwal (2013) proposed a model for predicting the credit risk from the merchants. This paper builds upon their technique to predict the fraud risk posed by the merchants to the banks. Fraud risk is an important aspect of risk management systems, particularly in the credit space. The uncertainty surrounding the receipt of paybacks calls for designing robust risk prediction models. Fraud risk is very different from credit risk because fraud risk does not follow a pattern. It happens suddenly, and may not always have a trend before it happens. This creates a need for separate model for fraud risk prediction. This paper develops a fraud risk prediction model that uses logistic regression technique, deployed using SAS. The setup of the study is the merchant-bank relationship in the credit card industry. The model developed in this paper triggers on a transaction level, and assigns a ‘probability score of default (PF) to each merchant for a possible fraud risk whenever a transaction is done at the merchant. Such a score warns the management in advance of probable future losses on merchant accounts. Banks can rank order merchants based on their PF score, and instead of working on the entire merchant portfolio, they can focus on the relatively riskier set of merchants. The PF model is validated by comparing the actual defaults with those predicted by the model and a good alignment is found between the two. The results show that the model can capture 62 percent frauds in the first decile when the transactions are sorted by the probability of fraud computed by the model.