Classification Using Optimization: Application to Credit Ratings of Bonds

Abstract

The classification approach, previously considered in credit card scoring, is extended to multi class classification in application to credit rating of bonds. The classification problem is formulated as minimization of a penalty constructed with quadratic separating functions. The optimization is reduced to a linear programming problem for finding optimal coefficients of the separating functions. Various model constraints are considered to adjust model flexibility and to avoid data overfitting. The classification procedure includes two phases. In phase one, the classification rules are developed based on “in-sample” dataset. In phase two, the classification rules are validated with “out-of-sample” dataset. The considered methodology has several advantages including simplicity in implementation and classification robustness. The algorithm can be applied to small and large datasets. Although the approach was validated with a finance application, it is quite general and can be applied in other engineering areas.