01 / THE DETAIL
The seven questions to answer
Document these items before choosing a method.
- Context: which part of the lending lifecycle needs improvement?
- Action: what will a person or policy do differently?
- Population: which borrowers, products and facilities are in scope?
- Target: what observable outcome are we trying to estimate?
- Clock: when is evidence available, and when is the outcome observed?
- Comparison: which rule, scenario or existing practice is the baseline?
- Responsibility: who reviews the evidence and owns the decision?
02 / THE DETAIL
Translate vague objectives
“Reduce risk” is not a target definition. It could mean fewer defaults, lower severity after default, less concentration or smaller cash-flow gaps. Each requires different evidence. Keep the business ambition, model target and evaluation measure as distinct parts of the brief.
03 / THE DETAIL
Describe the boundary of the answer
A model can estimate risk under a defined population and horizon. It cannot establish a causal uplift from a new price or collection action without suitable evidence. State what the model will not be used to decide, particularly for segments outside its evaluation scope.
EXAMPLE / PUTTING IT INTO PRACTICE
For a collections review, “estimate the chance of sustained cure within our defined horizon for eligible arrears episodes” is more actionable than “build an AI collections model”. The cure definition, eligibility and horizon still need to be supplied by the relevant operating policy.
WHAT YOU TAKE FORWARD