CMStatistics 2017: Start Registration
View Submission - CMStatistics
B0734
Title: Credit risk modelling via efficient network-based inference Authors:  Veronica Vinciotti - Brunel University London (United Kingdom) [presenting]
Elisa Tosetti - Brunel University London (United Kingdom)
Francesco Moscone - Brunel University London (United Kingdom)
Abstract: A credit risk model with network effects for a very large number of companies is developed. We assume a probit specification with group-specific effects having a non-diagonal, sparse covariance matrix, and adopt a penalised maximum likelihood estimation approach. Hence, we develop an Expectation-Maximization algorithm where we exploit the properties of truncated normals to proxy the conditional expectations. Monte Carlo experiments show that our proposed estimator has good finite sample properties and can be adopted for estimation and prediction using very large, or huge, datasets, given its moderate computation costs.We use a sample of nearly 568,000 accounts for Small and Medium-sized Enterprises (SMEs) in the United Kingdom in the years 2009 to 2013. We compare the prediction performance and estimated parameters of our credit risk model with that of a conventional default prediction model. We find that accounting for network effects makes a significant contribution to increasing the default prediction power of risk models built specifically for SMEs.