Bayesian clustering for portfolio credit risk

Credit risk models for loan portfolios typically assume that exposures can be assigned to homogeneous risk buckets, as in Vasicek-type and Basel-style portfolio credit risk models. We propose a Bayesian clustering model for constructing homogeneous risk buckets directly from loan credit histories. In contrast to traditional segmentation, the framework assigns weighted memberships across multiple clusters, capturing cross-sector and multi-factor exposures more realistically. Using both simulated and real credit data, we find that the proposed method can improve the estimation of loss distributions, Value-at-Risk, and Expected Shortfall. The approach provides a statistically robust and operationally tractable alternative to conventional bucketing, offering a more flexible foundation for portfolio credit risk management and capital assessment.