In the research work presented in this paper the possibility of improving scoring accuracy by pre-segmenting the population and then building subpopulation scorecards for each segment, is examined. A novel subspace selection technique together with a two-step cluster detection technique is used to identify naturally occurring segments (represented by clusters) in the data. Empirical results obtained from a real credit card data set demonstrate the strength of the approach. The segmentation method is able to separate the population into parts with widely different risk profiles (measured by the subpopulation odds). In the following step, different scorecards are built for each of these population segments. Significant improvement of overall accuracy of the scoring system is shown to be achieved by these segment-specific scorecards. Further, the scoring system resulting from the proposed approach is easy to implement in credit card account management systems that are currently used in the credit card industry.