10% reduction in loan losses through smarter credit scoring models

Financial Services, Risk Management

Our credit risk model used machine learning &  regression variants to reduce losses by 10%


  • Client was a large global bank
  • Indus re-developed the application stage risk model for Client’s Personal Loans business, followed by credit policy redesign
  • Indus tested various scorecards, horseraced multiple modeling techniques, validated on in & out of time samples, and developed final models
  • 1000+ predictor variables were used for modeling; Reject Inferencing done to calibrate model on turndowns


Multiple steps used to prepare the modeling datasets


Lift Chart for a sample segment



  • Tested multiple approaches – CART, CHAID, various regression techniques, and ensembling methods (bagging, boosting, random forests, etc)
  • Developed model suite led to 7 point Gini uplift vs. existing model; Bad rate 10% higher in worst deciles
  • Indus model has been deployed by the Client