The bank used quarterly scorecards that did not account for the dynamic behaviour of the customers. There were inaccuracies in thin-file and near-prime applicants’ classification. The credit department did not have visibility into the factors underlying lending decisions for each applicant.
We designed a new underwriting engine with the use of gradient-boosted models built based on transaction behaviour, payment history, and bureau data. Each decision became explainable with the factors used in scoring.
The underwriting engine included feature engineering, modeling, versioning, and deployment provided by Amazon SageMaker. The real-time inference was embedded in the loan origination process with visibility of factors for each decision made.
43% drop in default rate. Improvements in regulatory compliance due to explainability of each decision at the applicant level. Model updates transitioned from quarterly releases to a continuous pipeline in MLOps.
