A Critical Review of Machine Learning Approaches for Credit Risk and Loan Default Prediction

Approval or rejection of loans continues to be difficult decisions, and these decisions are becoming more difficult as financial markets continue to become globalised and interconnected. This review will look at research papers that examine how machine learning can assist with establishing more effective predictions of loan defaults in various lending environments, including U.S. traditional commercial banks (using both conventional and nonconventional forms of borrowing), P2P lending platforms (like Kiva and Lending Club), and mobile money networks in Sub-Saharan Africa. A common theme emerges across different studies referenced, which is that machine learning algorithms such as Gradient Boosting and XGBoost outperform traditional regression techniques (e.g., logistic regression) with a significant safety margin (generally 10%–15%), with 10% being the lowest margin of error for any study, and 30% being the highest margin of error across studies measuring AUC (area under the curve). The more significant findings are found when comparing average AUC measurement results to the AUC results of a dataset of behavioural signals (including factors related to how many times borrowers are late on payments, velocity of spends made by borrowers, and downloadable applications added to their mobile devices), all of these factors combined have produced much better predictive power than traditional financial ratios have produced on their own. When treating class imbalance - namely using methods such as SMOTE and ENN for resampling - recall has garnered considerable attention; similarly many of the major concerns are addressing issues including but not limited to: inequitable distributional burdens on various groups, the prevalence of model validation only using one dataset, the issue of temporal drift and the negative impact that changing economic conditions are having on how models perform. As part of this systematic review, previously published research will be systematically organized into categories according to their respective findings/gaps and ideas for areas of future research of the greatest potential.