MONOTONIC SUPPORT VECTOR MACHINES FOR CREDIT RISK RATING

Credit rating models are widely used by banking institutions to assess the creditworthiness of credit applicants and to estimate the probability of default. Several pattern classification algorithms are used for the development of such models. In contrast to other pattern classification tasks, however, credit rating models are not only expected to provide accurate predictions, but also to make clear economic sense. Within this context, the estimated probability of default is often required to be a monotone function of the independent variables. Most machine learning techniques do not take this requirement into account. In this paper, monotonicity hints are used to address this issue within the modeling framework of support vector machines (SVM), which have become increasingly popular in this field. Non-linear SVM credit rating models are developed with linear programming, taking into account the monotonicity requirement. The obtained results indicate that the introduction of monotonicity hints improves the predictive ability of the models.