This paper proposes a new fraud detection method using multiple classifiers to handle skewed data distributions. The method partitions the minority class using oversampling and trains Naive Bayes, C4.5, and backpropagation classifiers on each partition. The classifiers are then combined using stacking and bagging. Experimental results found that stacking and bagging the classifiers achieved higher cost savings than single classifiers, with stacking-bagging performing the best. Oversampling performed better than undersampling or SMOTE for this skewed fraud detection data.