Cost-sensitive learning for credit risk

Abstract

This thesis addresses the problem of fraud detection and credit risk from a cost sensitive perspective, exploring techniques that maximize the benefits to a financial institution while minimizing the probability of loss. First, an algorithm is proposed to estimate the optimal decision given a score. Once the optimal decision rule problem is solved, the estimation of the score is addressed from a cost-sensitiveperspective. A parametric model is proposed to estimate the score, and its consistency and asymptotic normality are obtained. A cost-sensitive semi-parametric model is also proposed, which is more robust and flexible. Finally, credit risk is approached from a reinforcement learning perspective. An online learning algorithm and a bandit algorithm are proposed to obtain updated models with the latest available information, optimizing decisions from a cost-sensitive perspective. The good performance of all proposals is demonstrated through extensive simulation studies and the analysis of various real credit risk data sets.