01 / THE DETAIL
Know what each record represents
State what one row represents: application, loan, instalment, payment, account-month or recovery event. Use stable identifiers to join only genuinely related records. Avoid duplicating a loan balance across payments and then adding it as if it represented additional exposure.
02 / THE DETAIL
Check the time and the outcome
Record both business dates and data availability where they differ. Separate predictor history from later outcomes. Keep the observation end for active accounts and incomplete workouts so their unobserved future is not treated as a known result.
03 / THE DETAIL
Review a practical inventory
Use this checklist to identify the records relevant to your business question.
- Originations: applications, decision dates, booked terms and later outcomes.
- Behaviour: schedules, receipts, balances and historical account states.
- Recovery: default exposure, dated proceeds, expenses and closure status.
- Growth: opening balances, funding maturities, cohort assumptions and constraints.
04 / THE DETAIL
Resolve gaps without inventing evidence
If units are unclear, obtain the source definition. If outcomes are immature, retain that status. If relevant history is absent, consider a transparent scenario or a narrower use. For decisions about your borrowers, check estimates against relevant loan outcomes. Where that history is missing, make the gap visible and use assumptions to explore possibilities rather than treating them as observed results.
EXAMPLE / PUTTING IT INTO PRACTICE
A recovery dataset omits accounts with no receipts. The apparent recovery experience would be overstated. Include the full eligible default population and distinguish zero receipts from missing records or ongoing recovery.
WHAT YOU TAKE FORWARD