Predicting bank loan recovery rates with neural networks

This study evaluates the performance of feed-forward neural networks to model and forecast recovery rates of defaulted bank loans. In order to guarantee that the predictions are mapped into the unit interval, the neural networks are implemented with a logistic activation function in the output neuron. The statistical relevance of explanatory variables is assessed using the bootstrap technique. The results indicate that the variables which the neural network models use to derive their output coincide to a great extent with those that are significant in parametric regression models. Out-of-sample estimates of prediction errors suggest that neural networks may have better predictive ability than parametric regression models, provided the number of observations is sufficiently large.