Design of Palatable Credit Scorecards as a Highly Automated Analytic Service by Combining Machine Learning with Domain Expertise

Lenders require accurate and interpretable credit scoring models palatable to regulators, financial services staff and consumers. Expert-designed segmented scorecards fill this need. Building such models is a laborious data-guided task for experienced modelers. It can take weeks to hone a model for deployment. Lenders would like to design, update and test models, predictors and segmentation schemes more frequently, objectively and cost-effectively, as environments change fast and as new data emerge. We propose scorecard design as an automated analytic computing service used by domain experts, comprising data-driven machine learning with expert-imposed palatability restrictions and model visualization, in two stages: Stage I fits a tree ensemble model to render a best-fit score and a list of segmentation candidates. Stage II uses this information to generate optimal palatable segmented scorecards subject to restrictions provided by the experts. When implemented on a computer cluster, our process yields close to deployment-ready scorecards within minutes to hours, which can be rapidly honed and upon approval deployed into a separate scoring service. While motivated by transparency needs of credit scoring, such a service can be valuable for any application requiring highly predictive yet palatable scoring algorithms.