Files
Abstract
Financial institutions need to measure risks within their credit portfolios (homeloans, credit card, auto loans, etc.) for regulatory requirements and for internal riskmanagement. To meet these requirements financial institutions increasingly rely onmodels and algorithms to predict losses resulting from customers' defaults. Hence,developing sufficiently accurate and robust models is one of the major efforts of quantitativerisk management groups within these institutions.The proposed research is one such effort to develop robust and efficient models forthe credit default risk problem. Specifically, the research focuses on developing a logisticregression based model and machine-learning based non-parametric algorithmsto predict default risk for credit card accounts. We use data named "default of creditcard clients data set" sourced from the University of California, Irvine to pursue thisstudy.The thesis provides a systematic step-by-step model development approach - startingfrom building a benchmark model, continually improving the benchmark modelby tuning the hyper-parameters, recursively eliminating insignificant variables fromthe predictor pool, and then evaluating model on various performance measures - toarrive at the best estimation for the model to implement on the training data.Finally, based on the research findings, we provide insight into opportunities, andfuture research in using machine-learning based modeling approaches for addressingcredit default risk.