Abstract

Risk assessment of credit portfolios is essential in banking. The bank with the most accurate view on its credit risk, will be most profitable. In order to calculate the risk, each client's 'probability of default' needs to be estimated. The probability of default is defined as the probability that a client can not meet its repayment obligations toward the bank anytime in the next twelve months. Credit rating models assign a probability of default to each client, based on a set of input variables. In this research, the best practice modeling method, logistic regression, is benchmarked with Bayesian networks. A Bayesian network is a graphical representation of a probabilistic model. Di#erent Bayesian network structures return different results, ranging from bad to good. Some structures require advanced learning techniques. Opportunities to improve these techniques are proposed. This research was performed at ABN AMRO Group Risk Management and in the Decision Systems Laboratory of the University of Pittsburgh.