Application of machine learning algorithms in internal rate based probability of default models

Economic turmoil that affects the financial system, results with severe and long-termed effects globally as it can also be seen in the past. As one of the most important elements in the financial system, banks always need to be prepared for the worsening macro-economic conditions to maintain financial stability and prevent a too big to fail event. Capital structure and the capital amount that banks need to have, play crucial roles to prepare them for such events. In the early 1980s, the Basel rules are set by the Basel Committee which is gathered by Bank of International Settlement (BIS) to maintain a strong capital structure for the banks. As a consequence of the Basel 2, banks have the opportunity to calculate the risk parameters such as Probability of Default (PD), Loss Given Default (LGD), and Exposure at Default (EAD) internally to calculate their Risk-Weighted Assets which will be further used to determine the amount of capital that they need to set aside according to the Capital Adequacy Ratio which is one of the Basel requirements. These parameters are calculated via internal models of the banks (IRB Models) by using predictive modeling. In this study, different machine learning models have been developed and then further assessed both quantitatively and qualitatively with respect to their output performance and their eligibility to be used as IRB PD model respectively, and then further challenged against the current practice which is Logistic Regression. Key Words: Credit Risk, BASEL, Machine Learning, Internal Rating Based Approach