This paper proposes a new loss function for unbalanced binary data sets with overlapping classes. This loss function is a weighted cross-entropy binary function that takes as its argument the distance between an observation of the majority class and an observation of the nearest minority class. The distance is calculated using the nearest neighbor algorithm. The quality of this loss function is measured with the AUC-ROC metric on a fraud detection data set for energy saving certificates. In addition, the reduction in the workload for human verification is calculated. The new loss function improves the result. A statistical test confirms the improvement of the model using the new loss function.