Challenge

A near-prime lender had a revenue opportunity locked behind conservative linear scorecards. Their textbook logistic regression models rejected creditworthy applicants who didn't fit conventional credit profiles. The business case was clear; every basis point of approval rate improvement translated directly to bottom-line revenue.

Rebuilding and redeploying scorecards took 3 months, making it impossible to adapt to post-pandemic shifts in consumer spending behaviour or changing credit bureau dynamics. Each model development cycle required extensive manual feature engineering, separate validation testing, and lengthy approval processes.

"We knew conservative scorecards were costing us revenue, but we couldn't prove how much or move fast enough to capture the opportunity. Model development timelines meant decisions made in Q1 deployed in Q4—by which time market conditions had shifted again." CRO

Goals

• Deploy explainable ML scorecards achieving materially higher calibration than incumbent bureau score-based models while maintaining full regulatory transparency
• Implement a segmented modelling strategy with differentiated models for each product and client type, optimising risk discrimination within each population.
• Build elastic MLOps infrastructure, reducing model development cycles from 12+ weeks to under 6 weeks, enabling continuous improvement and rapid response to portfolio drift.
• Establish a production Model Risk Management lifecycle with automated champion-challenger testing, Population Stability Index monitoring, and one-click model replacement capabilities.

Solution

We rebuilt credit decisioning on DataRobot, testing over 100 machine learning algorithms in three weeks rather than building two models over three months. We deployed Generalised Additive Models (GAMS) because they solve the interpretability-performance tradeoff that constrains financial services. We implemented a comprehensive reject inference methodology to address sample selection bias.

We rebuilt the credit card underwriting system on DataRobot, changing how scorecards get developed and deployed. Instead of building a handful of models over several months, we tested over 100 machine learning models in three weeks. This rapid experimentation identified which algorithms delivered the best risk discrimination for their specific portfolio mix.

We deployed Generalised Additive Models because they explain predictions (model explainability) while outperforming legacy linear models (model accuracy). The architecture combines traditional credit bureau data with behavioural patterns to assess near-prime applicants who don't fit conventional lending profiles. We implemented Model Risk Management lifecycle processes that handle model versioning, champion-challenger testing, and automated retraining. The infrastructure enables one-click model replacement and real-time performance monitoring.

Development time dropped from 3 months to 6 weeks because the platform automates feature engineering, model training, and validation testing. Risk teams can now test new approaches without disrupting production operations and track model performance against shifting market conditions.

Impact

Approval rates increased 10%+ while reducing bad rates. This directly translates to revenue gains, thousands of additional customers accessing credit products without weakening portfolio quality. The improved risk discrimination came from models that assess patterns traditional scorecards miss, particularly for near-prime applicants with non-standard financial profiles.

Model development velocity changed how the organisation responds to market conditions. 6-week cycles instead of 3-month cycles mean they can adapt to regulatory changes and consumer behaviour shifts 4X times faster. The MLOps infrastructure reduced the total cost of ownership for the underwriting platform through automated retraining and deployment pipelines.

The system provides real-time model performance tracking, automated compliance documentation, and champion-challenger testing that runs continuously in production. These capabilities mean improvements compound over time rather than requiring periodic rebuilds.