Define what the score predicts
A score is useful only when its target and time horizon are clear. Predicting a missed payment in 30 days differs from predicting recovery within 90 days.
Also document scale direction. A higher value may mean greater risk in one model and better repayment prospects in another. A rank of 80 on a 100-point scale is not automatically an 80% probability.
Inputs and interpretation
Models may use eligible payment history, days past due, balances, and account events. Inputs should be relevant, accurate, permitted for the intended use, and available at the time of prediction.
A dispute is an account issue requiring appropriate handling; it should not automatically be treated as unwillingness to pay. Likewise, a score does not establish fraud or ability to pay.
Example
Two accounts have the same balance and delinquency stage. A validated model may estimate different near-term repayment outcomes from their payment histories. An operational priority decision must still consider restrictions, customer requests, cost, and the approved strategy.
Governance and monitoring
Document the target, model version, data freshness, validation population, and permitted actions. Compare predictions with observed outcomes and review errors across relevant segments. Monitor drift and provide a route for correction or human review.
The NIST AI Risk Management Framework provides voluntary guidance on reliability, transparency, privacy, and harmful bias.
Related terms: High-Risk Accounts and Aging Buckets.