Credit risk modeling using a weighted support vector machine

Abstract

The modeling of credit risk is traditionally based on approaches such as linear regres- sion or multiple discriminant analysis. There are several limitations to these methods, in particular their inability to adapt to new data and the assumption of a certain (linear) relation between the dependent and independent variables. In this thesis we will use a new machine learning technique called weighted support vector machine combined with averaged stochastic gradient descent. Using this approach, we create a classification of a data set containing mortgage loans of Freddie Mac into several groups with increasing probabilities of default. This method shows promising results, both in terms of predictive and discriminatory power, especially when information about the monthly performance of these loans and the macro-economic situation is included. If the performance on other data sets is similar, the technique can be implemented for credit risk modeling.