More accurate credit origination with segment-based models
Banking
Business
Corporate
+ 2.000 employees
Type of solution
Redesign of origination scoring models
Segments
evaluated based on each client’s available information
Challenge
A leading bank had outdated origination models for credit cards and consumer loans. They were applied to all prospects and applicants, even segments with little data support, and used many variables with no intuitive link to risk.
Solution
We refined the default definition, excluding inconclusive or non-credit-risk events, and designed models based on each client’s history. We compared logistic regression and gradient boosting on predictive power and business metrics.
Technologies used
Databricks · Logistic Reg. (WOE) · LightGBM · MLflow
Impact
Segment-based models improved performance and set apart clients without enough data to be evaluated, instead of assigning them an unreliable risk. Best practices for monitoring and retraining were put in place.
Related services
The capabilities that made this project possible.
