Comparison of ensemble nlearning and neural network models for default risk analysis: a reproduction study

Researcher: Tebogo Malatsi, University of the Witwatersrand, Johannesburg
Supervisor: Assoc. Prof. YudhvirSeetharam, University of the Witwatersrand, Johannesburg

This study replicates prior research on default risk assessment using ensemble and deep learning techniques, leveraging payment data from the UCI Machine Learning Repository for Taiwan. It specifically compares boosting, random forest, and neural network models with Tanh and ReLUactivations, evaluating their predictive accuracy and classification capabilities through metrics like accuracy, AUC, ROC curve, and F-Score.
The study reveals trade-offs in model performance metrics, with heightened accuracy but nuanced reductions in AUC/F-Score. Random Forests demonstrate superior learning and generalization, while neural network models emphasize the challenge of balancing sensitivity, specificity, precision, and recall.