Assessing likelihood of repayment.
WHAT THE TARGET IS
Default, defined by a period and a threshold.
WHAT MUST BE DEFINED
How many days overdue counts Over what period after lending
WHAT SELECTION BIAS MEANS HERE
You only observe outcomes for applicants who were accepted.
WHY THAT IS THE CENTRAL DIFFICULTY
The model learns from a population filtered by previous decisions.
WHAT REJECT INFERENCE ATTEMPTS
Estimating outcomes for rejected applicants.
WHAT TO ACCEPT
That it is approximate, and conclusions should be cautious.
WHAT REGULATION FREQUENTLY REQUIRES
Explanation of adverse decisions Absence of prohibited factors Documented governance
WHAT THAT ARGUES FOR
Interpretable models, or interpretable approximations with careful validation.
WHAT SCORECARDS PROVIDE
Points per attribute, transparent and auditable.
WHY THEY PERSIST
They are explainable, stable, and defensible.
WHAT TO MONITOR
Score distribution shift Approval and default rates by segment Performance as economic conditions change
WHY THAT LAST POINT MATTERS
Models trained in benign conditions perform differently in a downturn.
WHAT OBLIGATIONS MAY APPLY
Lending and consumer credit regulation. Take advice on your position.
