We discuss feature selection approaches within the context of linear programming models for discrimination, with special emphasis on a new interpretation and application of the jackknife. We propose new criteria for variable selection based on viewing the jackknife as a heuristic for a bicriterion problem. Furthermore, instead of using the jackknife estimates for constructing the discriminant, we obtain it from an independent run of the linear program. In the context of benchmarking our approach, we discuss alternative accuracy criteria and propose a new one driven by decision-making practice. Several implementation refinements are also proposed, which appear to improve predictive accuracy. Experiments on two large real-life databases of credit applicants, one of mortgages and one of auto loans, illustrate the advantages of our approach.