Most approaches to credit scoring generate model parameters by minimising some function of individual error, or by maximising likelihood. In practice, the criteria by which the parameters of a model are determined and the criteria by which models are assessed may differ. Practitioners tend not to be interested in standard measures such as the R2 coefficient for linear regression or the likelihood ratio for logistic regression. Instead, performance will be assessed using global measures such as the GINI coefficient, or by considering the misclassification rate at different points in the distribution of model scores. In this paper an approach using genetic algorithms is described, where the training algorithm is used to directly maximise/minimise the performance measure of interest. Empirical results are presented, showing that genetic algorithms have the potential to generate scoring models that are competitive with models constructed using more traditional approaches, and that there is scope for improved models when prior information about model usage is incorporated within the parameter estimation process.