Estimating default probabilities of CMBS loans with clustering and heavy censoring

This paper provides a comprehensive default estimation of commercial real estate loans with a complete commercial mortgage backed securities (CMBS) loan history database. Standard survival models assume that eventually every observation will experience the event. However, often there is a high proportion of censored observation in the sample. A mixture model is proposed to disentangle the probability of "long-term survivorship" and the timing of default occurrence. Loans within the same geographical area and property type tend to exhibit correlation in default incidence. A multilevel model is proposed to capture this correlation within and between clusters.