An Investigation into the Use of Generalized Additive Neural Networks in Credit Scoring

Logistic regression occupies a central position in the field of credit scoring as it is relatively well understood and an explicit formula can be derived on which credit decisions may be based. Although artificial neural networks may be more powerful than logistic regression, it is not widely used in credit scoring because it is a black box with respect to interpretation and the absence of reasons why the neural network has reached its decisions may be unacceptable. In contrast to logistic regression and neural networks, generalized additive models is a compromise between inflexible, but docile linear models and flexible, but troublesome, universal approximators. In this study the performance of a generalized additive model (implemented as a neural network and therefore called a generalized additive neural network or GANN) is compared to that of a logistic regression model on a home equity data set, where the aim is to predict whether an applicant will eventually default or be seriously delinquent on a loan. Partial residuals are used to investigate the effect of the individual inputs, adjusted for the effect of the other inputs, where the jth partial residual is the deviation between the actual values and that portion of the fitted model that does not involve variable xj. Using a GANN architecture therefore assists in alleviating the black box perception of neural networks with respect to interpretation as the effect of each input variable on the fitted model can be interpreted using a graphical method (partial residual plot). Another benefit of using a GANN is that the nonlinear effects of the inputs may be easier to learn in this constrained architecture than in the general artificial neural network or multiplayer perceptron setting.