AI Collection Scorecard: Predict Repayment Likelihood

Debt collection specialists in the analytical field play a vital role in predicting the likelihood of recovery and optimizing collection strategies. One useful tool for achieving this is a collections scorecard. In this article, we will outline how to build a comprehensive collections scorecard that accurately predicts the probability of recovery.

I. Variables and Data Types

A. Collection Behavior Data

1. Number of phone calls made in the last 1, 2, and 3 periods
2. Total number of phone calls made
3. Number of phone calls connected in the last 1, 2, and 3 periods
4. Total number of phone calls connected
5. Connection rate for the last 1, 2, and 3 periods
6. Pverall phone connection rate
7. Number of days lost contact in the last 1, 2, and 3 periods
8. Total days lost contact
9. Number of valid friends
10. Number of valid contacts
11. Average call duration per call
12. Total call duration
13. Total number of call backs
14. Total collection messages sent
15. Total number of collectors involved
16. Total number of collection letters sent

B. Customer Personal Information

1. Age
2. Gender
3. Occupation
4. Education
5. Monthly income
6. Marital status
7. Housing situation

Frequently asked questions

What repayment outcome and prediction period should the scorecard use?

Start with an outcome your collections team can observe and act on, such as whether an account makes a payment within 30 days of scoring. Choose a period that matches how often you assign accounts and review results. Define in advance whether partial payments, settlements, and promises kept count as success, or the scorecard may rank accounts against an outcome that does not match your collections goal.

Where should we set the cutoff for assigning higher collection effort?

Set the cutoff from the expected recovery value and the cost and capacity of the additional contact, not from a generic score threshold. Test a few score bands in a controlled pilot, then compare collected amounts and contact costs for each band. A high repayment likelihood does not always mean an account needs more effort, so check whether extra contact changes outcomes rather than simply identifying people who would have paid anyway.

How should we test whether call history improves repayment predictions?

Use a time-based holdout: build the scorecard on earlier accounts and test it on later accounts that were not used to fit it. Include only call and contact information available at the moment each account is scored. Contact history can reflect earlier collection choices as much as a person’s willingness to repay, so compare performance across existing contact policies before using those fields to allocate more effort.

When should we refresh the collections scorecard?

Review performance at least quarterly, and sooner if repayment patterns, products, or collection policies change materially. Track whether predicted repayment rates match observed rates across score bands, and whether the ranking still separates accounts that repay from those that do not. Rebuild only when the review shows meaningful deterioration, then validate the replacement on later accounts before putting it into use.

Put these checks into a decision flow

Create rules for credit, fraud or collections, then run them on live decisions.

Start free See pricing