Consistency of Banks' Internal Probability of Default Estimates

Some financial institutions can use internally developed credit risk models to determine their capital requirements. At the same time, the regulatory framework governing such models allows institutions to implement diverse rating systems with no specified penalty for poor model performance. To what extent the resulting model risk { potential for equivalent models to deliver inconsistent outcomes { is prevalent in the economy is largely unknown. We use a unique dataset of 4.9 million probability of default estimates provided by 28 global IRB banks, covering the January 2016 to June 2020 period, to assess the degree of variance in credit risk estimates provided by multiple banks for a single entity. In line with the prior literature, we find that there is a substantial variance in outcomes and that it decreases with the amount of available information about the assessed entity. However, we further show that the level of variance is highly dependent on the entity type, its industry and locations of the entity and contributing banks; banks report a higher deviation from the mean credit risk for foreign entities. Further, we conclude that a considerable part of the variance is systematic, especially for fund models. Finally, utilising the latest available data, we show the massive impact of the COVID-19 pandemic on dispersion of credit estimates.