01 / PUT IT INTO PRACTICE
Begin with the records the business already keeps
Different decisions need different histories. Application records explain what was known before lending. Schedules and payments explain behaviour. Recoveries and expenses explain losses after default. A single current loan-book export rarely contains all four.
- Loan master: identifiers, product, origination date, contractual terms and security.
- Transactions: dated amounts, allocations, corrections and currency.
- Events: defaults, cures, restructures, closures and observation end.
- Portfolio context: funding, costs, capital and scenario assumptions where needed.
02 / PUT IT INTO PRACTICE
Map meaning before matching column names
A column called “balance” could mean principal only or a wider claim. A payment date could be receipt date or posting date. Define units, row grain, identifiers, permitted values and missingness explicitly. Preserve raw records so a derived feature can be traced back to its source.
03 / PUT IT INTO PRACTICE
Separate missing, zero and not yet observed
No recorded recovery is not always a completed zero-recovery outcome. A new loan may not have matured through the target horizon. A closed account is different from one that simply disappeared from the export. These distinctions affect both fitting and evaluation.
04 / PUT IT INTO PRACTICE
Let suitability determine the next step
Assess event coverage, observation length, segment consistency, effective sample and uncertainty. There is no universal row count that makes a model suitable. The outcome may be ready for a defined evaluation, limited to scenario analysis, or not supported for the intended use.
EXAMPLE / PUTTING IT INTO PRACTICE
An export has current balances but no historical due dates or receipts. It can describe today’s portfolio, but it cannot establish a repayment trend. The next request is for dated schedule and payment records, not a guessed “behaviour score”.
WHAT YOU TAKE FORWARD