The problem of unpaid bank debts is becoming increasingly important in developed countries. Many empirical works are being published in an attempt to find a model capable of determining as accurately as possible whether an individual requesting a loan will be able to pay it back. This paper analyses the predicting capability of one non-parametric and two parametric models. As regards the former, the often-overlooked problem of overlearning is also tackled using the cross-validation technique. Furthermore, a three-level grading of loan applications is proposed depending on their likely performance: grant, refuse, or doubtful hence subject to manual consideration by bank staff.
Credit scoring; logit; discriminant analysis; classification trees; cross-validation