No abstract

No abstract

For banks, credit lines play an important role exposing both liquidity and credit risk. In the advanced internal ratings-based approach, banks are obliged to use their own estimates of exposure at default using credit conversion factors. For volatile segments, additional downturn estimates are required. Using the world's largest database of defaulted credit lines from the US and Europe and macroeconomic variables, we apply a Bayesian mixed effect quantile regression and find strongly varying covariate effects over the whole conditional distribution of credit conversion factors and especially between United States and Europe. If macroeconomic variables do not provide adequate downturn estimates, the model is enhanced by random effects. Results from European credit lines suggest that high conversion factors are driven by random effects rather than observable covariates. We further show that the impact of the economic surrounding highly depends on the level of utilization one year prior default, suggesting that credit lines with high drawdown potential are most affected by economic downturns and hence bear the highest risk in crisis periods.

The calibration of financial models is laborious, time-consuming and expensive, and needs to be performed frequently by financial institutions. Recently, the application of artificial neural networks (ANNs) for model calibration has gained interest. This paper provides the first comprehensive empirical study on the application of ANNs for calibration based on observed market data. We benchmark the performance of the ANN approach against a real-life calibration framework that is in action at a large financial institution. The ANN based calibration framework shows competitive calibration results, roughly four times faster with less computational efforts. Besides speed and efficiency, the resulting model parameters are found to be more stable over time, enabling more reliable risk reports and business decisions. Furthermore, the calibration framework involves multiple validation steps to counteract regulatory concerns regarding its practical application.

The calibration of financial models is laborious, time-consuming and expensive, and needs to be performed frequently by financial institutions. Recently, the application of artificial neural networks (ANNs) for model calibration has gained interest. This paper provides the first comprehensive empirical study on the application of ANNs for calibration based on observed market data. We benchmark the performance of the ANN approach against a real-life calibration framework that is in action at a large financial institution. The ANN based calibration framework shows competitive calibration results, roughly four times faster with less computational efforts. Besides speed and efficiency, the resulting model parameters are found to be more stable over time, enabling more reliable risk reports and business decisions. Furthermore, the calibration framework involves multiple validation steps to counteract regulatory concerns regarding its practical application.