Genetic algorithms for credit scoring: Alternative fitness function performance comparison

•Performance of three different fitness functions in credit scoring is evaluated.•A novel fitness function based on subsets of client characteristics is proposed.•The proposed fitness function shows better accuracy and sensitivity. Credit scoring methods have been widely investigated by researchers; recently, genetic algorithms have attracted particular attention. Many research papers comparing the performance of genetic algorithms and traditional scoring techniques have been published, but most do not provide enough detail about the fitness function used by the genetic algorithm—despite the fact that fitness function has a key influence on the model’s overall performance. The aim of this paper is to evaluate the predictive performance of different fitness functions used by genetic algorithms in credit scoring. An alternative fitness function based on a variable bitmask is proposed, and its performance then compared with fitness functions based on a polynomial equation as well as an estimation of parameter range. The results suggest that the bitmask is superior to the two other methods in both accuracy and sensitivity. The Wilcoxon matched-pairs sign rank test and paired t-Test indicate these results are statistically significant.