Ensemble Machine Learning for Credit Risk Prediction in Vietnam: Evidence from Loan-Level Bank Data

Abstract:

The two types of classification error in credit scoring carry very different costs: a false negative is far more expensive than a false positive. Using 57,187 loan-level records from a Vietnamese commercial bank, of which 1.23% are non-performing, we evaluate six machine learning models and four voting ensembles. Because missing a non-performing case is the costlier error, we prioritise recall and the F2 score in model selection. We also separate application-time predictors from post-origination information, evaluating each model against only the data available at its decision stage. The behavioural model reaches an AUC of 0.992, reflecting concurrent discrimination driven by accumulated loan tenure rather than early warning. Restricted to origination information, discrimination falls to 0.82-0.87, yet the model still ranks risk usefully, with the riskiest fifth of applicants containing about 72% of non-performing cases. Repayment- related variables provide the most predictive information, and recall-oriented ensembles catch most non-performing borrowers, with recall between 0.91 and 0.93. The practical message is that model evaluation should be tied to the information available at each decision point and to the relative cost of the errors that matter most.