infonet-economy

For creating or adjusting credit scoring rules, usually only the accepted applicant's data and default information are available. The missing information for the rejected applicants and the sorting mechanism of the preceding scoring can lead to a sample selection bias. In other words, mostly inferior classification results are achieved if these new rules are applied to the whole population of applicants. Methods for coping with this problem are known by the term "reject inference." These techniques attempt to get additional data for the rejected applicants or try to infer the missing information. We apply some of these reject inference methods as well as two extensions to a simulated and a real data set in order to test the adequacy of different approaches. The suggested extensions are an improvement in comparison to the known techniques. Furthermore, the size of the sample selection effect and its influencing factors are examined.