Core Path · L6.3.03

Building a scorecard

14 minScorecard pointsReject inferenceCut-off strategySwap setsBehavioural scoringFair treatment

Know these first

By the end you will be able to

  1. 1Turn a logistic model on weight-of-evidence inputs into scorecard points on a fixed scale.
  2. 2Explain why a scorecard built on accepted loans misjudges the declined, and what reject inference can and cannot fix.
  3. 3Set a cut-off from break-even odds and profit, and read a swap set.
  4. 4Describe a behavioural scorecard, and check a scorecard's outcomes across age groups against UK fair-treatment rules.

Why it matters

L3.3.14: Retail credit portfolios scored RiskAtoms Bank plc's personal loans on a booked scale: 600 points at odds of 50 to 1, 20 points to double the odds. The model of L6.3.02: Logistic regression for default can be printed on the same scale as a table of points. Set it to approve as many applicants as the old scorecard did and it swaps 1,379 of them for 1,379 others. In the simulation, where every applicant's outcome is known, the newcomers go bad at 9.0% and the ones it turns away at 8.8%. A scorecard is more than a fit; this lesson builds one and asks what it really changes.