Logit Regression in Forecasting the Insolvency Risk of the Customers of Automotive Financial Services

In this paper, in order to forecast the performance of instalments payments of the customers of an automotive financial service, we shall resort to a logit and a tree multivariate regression model. Our data set consists of 298,902 customers of an automotive financial service, whose application has already been successfully scored, and it includes 33 variables arising from merged different data sources. Some of these variables will be used to construct our target variable which enables us to discriminate between good and bad customers; in this way we observe in our data set a rate of bad customers of 6.77%. We first focus our attention to the problem of the right choice of the variables to be plugged into both the models, showing how a bad choice, which can arise resorting to automatic variables selection methods, can highly distort the forecasting power of the models. Furthermore a validation of the chosen set of the explanatory variables will be given by mean of a confirmatory Factorial Analysis. We finally turn to the choice of the best model and to the check of its power; this will be achieved on the basis of the results we obtain testing both the models on a randomly selected control group, which is hold out of sample.