Financial institutions usually adopt default prediction models to evaluate loan applicants, in order to distinguish those who are expected to pay back their debt from those who are likely to default. Although default prediction models cannot perfectly separate the applicants who will fully repay the loan from defaulters, they can significantly improve the allocation of financial resources, enabling lenders to grant credit to borrowers who could potentially have been excluded despite being creditworthy. We compare the parameters estimated from the bivariate probit model with sample selection with a standard probit model, using a training sample of 2,272 observations randomly drawn from the whole sample of 3,441 approved loans. Given the weak correlation between the unobservables of the selection and the outcome equation, the estimates of both models are very similar and consequently they have similar predictive performance. In line with Boyes et al. (1989), we find that a set of financial variables that increase (decrease) the probability of positive granting decision do not reduce (raise) the likelihood of a default. On the other hand, dummy variables describing the destination of a loan have a lower probability of being accepted as well as of turning into bad loans, compared to a more general category of loans without a precise purpose.
Financial intermediation and guarantee-backed loans: an analysis of default
VEZZULLI, ANDREA
2008
Abstract
Financial institutions usually adopt default prediction models to evaluate loan applicants, in order to distinguish those who are expected to pay back their debt from those who are likely to default. Although default prediction models cannot perfectly separate the applicants who will fully repay the loan from defaulters, they can significantly improve the allocation of financial resources, enabling lenders to grant credit to borrowers who could potentially have been excluded despite being creditworthy. We compare the parameters estimated from the bivariate probit model with sample selection with a standard probit model, using a training sample of 2,272 observations randomly drawn from the whole sample of 3,441 approved loans. Given the weak correlation between the unobservables of the selection and the outcome equation, the estimates of both models are very similar and consequently they have similar predictive performance. In line with Boyes et al. (1989), we find that a set of financial variables that increase (decrease) the probability of positive granting decision do not reduce (raise) the likelihood of a default. On the other hand, dummy variables describing the destination of a loan have a lower probability of being accepted as well as of turning into bad loans, compared to a more general category of loans without a precise purpose.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.