B1049
Title: Credit risk model with network effects for a large panel of companies
Authors: Veronica Vinciotti - Brunel University London (United Kingdom) [presenting]
Elisa Tosetti - Brunel University London (United Kingdom)
Francesco Moscone - Brunel University London and Ca Foscari of Venice (United Kingdom)
Abstract: A credit risk model is developed with network effects for a very large panel of companies. We assume a probit specification with group random effects having a non-diagonal, sparse covariance matrix. We propose a penalised maximum likelihood estimation approach and develop an Expectation-Maximization algorithm where we exploit the properties of truncated Normals to efficiently approximate conditional expectations. A simulation study shows good properties of our approach at a significantly reduction in computational cost. We fit this model on a dataset of over one million accounts for small and medium-sized enterprises in the United Kingdom over the period from 2009 to 2015. We group companies according to their sector and geographical location. We find that accounting for network effects makes a significant contribution to increasing the default prediction power of risk models built specifically for SMEs, compared to a more conventional probit prediction model. Our results may help bankers to improve their credit scoring of SMEs ultimately reducing their propensity to apply excessive lending restrictions.
Journal