Creating sophisticated credit scoring models is a useful method of locating defaulters. Extensive study was done on the ensemble credit rating since the ensemble learning approaches proven to be more effective than the individual classifiers. There are no standardised benchmark features for bank clients' credit scores, and each country and bank decides which standards to use based on the information it has available about its customers. Therefore, using the same classifiers on several data sets may yield excellent results for certain data sets but not others. Combining the best classifiers from among those available for each dataset independently with the ability to diversify the classifiers in ensemble learning methods produces excellent results for all datasets. In this work, we developed 6 different multilevel stacking method in which we carried out the credit scoring using 3 base classifiers at level 0, 3 meta-classifiers at level 1 and MLP as final classifier at level 2. The German and Australian UCI data sets were used to evaluate the suggested methodology. The ensemble models that combined KNN+DF+NB in level-1 and SVM+LR+RF in level-2 and MLP as final meta-learner at level 3 exhibited the best accuracy, AUC, and F1-score performance. The findings of this study showed that both balanced and unbalanced data sets produce highly favourable results using the suggested methodology