Profit scoring for credit unions using the multilayer perceptron, XGBoost and TabNet algorithms: Evidence from Peru

Abstract

Credit unions are growing microfinance institutions that base their lending decisions on the judgment of their credit analysts. Therefore, the purpose of this paper is to design 6 profit scoring models, capable of predicting the Internal Rate of Return (IRR) of credit applications, using the multilayer perceptron, XGBoost and TabNet algorithms and thus serve as a support tool for the credit analyst. For this purpose, the least correlated and most independent features were selected from the dataset coming from a Peruvian credit union and composed of 36 402 observations. Then, the hyperparameters of all algorithms were tuned. Finally, the profit scoring models that considered only the selected features were compared to which considered all features. As results, it was obtained that the most significant features that determine the IRR of a loan are the effective monthly interest rate and the member's maximum or average days delinquent. The results obtained from the performance evaluation of the profit scoring models suggested the XGBoost as the best algorithm. In addition, the model that used the XGBoost algorithm and considered all the features had the best performance.