projects / 2026 / Built at work · pilot
Collections Prioritization model
An XGBoost model that ranks overdue loan accounts by recovery likelihood, with an interpretable scorecard benchmark and LLM-written segment reports.
What it is
A model that ranks overdue accounts by how likely they are to be recovered, so collections teams know where to spend their time first. It's live for a pilot group in real-user testing ahead of a wider rollout at AYE Finance.
How it works
- Ranking model: XGBoost, trained on account and repayment history.
- Interpretable benchmark: a Logistic Regression scorecard with WOE/IV binning, built alongside it so the gradient-boosted model always has a transparent baseline to beat. In a regulated lender that comparison matters as much as the accuracy number.
- Imbalance: handled with SMOTE and cost-sensitive thresholding, which notably improved minority-class recall.
- Monitoring: PSI-based drift monitoring on input features, with MLflow for model versioning.
- Reporting: an LLM-generated layer that narrates segment trends in plain language for the people using the ranking.