An Algorithmic Model for Retail Credit Portfolio Segmentation

Under the new Basel bank capital framework, a bank must group its retail exposures into multiple segments with homogeneous risk characteristics. The U.S. regulatory agencies believe that a bank may use the internal models, including the loan-level risk parameter estimates such as PD and LGD, to group exposures into the resultant segments with homogeneous risk attributes. In contrast to the conventional decision tree method, we propose a new algorithmic technique for retail consumer loan portfolio segmentation. This new technique identifies the optimal number of segments, sorts the individual loan exposures into the various segments, and then leads to a greater degree of risk homogeneity in comparison to the baseline equal-bin and quantile-bin schemes. Furthermore, we analyze the Monte Carlo implied asset correlation values for the retail loan segments over time to help assess the implications for bank capital measurement. Our recommended method for retail credit portfolio segmentation results in some capital relief that serves as an incentive for the bank to invest in this alternative segmentation. This positive outcome accords with the core principle of statistical conservatism that is enshrined in the Basel regulatory requirements for bank capital measurement