The business problem
A card or personal-loan book usually starts collecting once a payment is already late. By then the account is in the 1 to 30 days past due bucket, the call centre is working from a list sorted by balance or days overdue, and the customers who were about to fall behind for the first time get the same letter as everyone else. Lenders that move contact earlier, before the due date, with digital reminders for low-risk customers and people on the phones for the high-risk ones, report fewer accounts rolling into later buckets with the same headcount.
I wanted to see how much of next month’s missed payments a lender can reach before the due date by ranking every account on its recent repayment and balance behaviour, and placing each one in a call, reminder or routine lane sized to the team.
- Team
- Head of collections, early-arrears team leaders, hardship specialists, credit risk analysts
- Book
- Hundreds of thousands of card or loan accounts
- Systems
- Card or loan platform export, Salesforce or a collections system such as Debt Manager, dialler and SMS
- Constraint
- National Credit Code hardship obligations; scores order calls and never decide credit
What was built
A monthly scoring step that runs when statements close. It reads the last six months of bills, payments and late flags for each account, scores the risk of missing next month's payment, and puts every account in one of three lanes. The lanes are written to the CRM as a call list with reason codes, and a team lead reviews the list before calls start.
| Lane | Accounts | Went on to miss | Share of all misses |
|---|---|---|---|
| Call this week (top 20%) | 1,200 | 55% | 50% |
| Reminder (next 30%) | 1,800 | 22% | 29% |
| Routine (bottom 50%) | 3,000 | 9% | 21% |
Lanes cut at score tenths on 6,000 held-out accounts. The actions are how a team would use each lane; their effect was not tested here.
Results
Cheaper than the rule at the same team size
| Method | AUC (95% CI) | PR-AUC | Brier |
|---|---|---|---|
| Rule: late last month | 0.672 (0.658–0.687) | 0.351 | 0.228 |
| Logistic (PAY_1 only) | 0.676 (0.659–0.695) | 0.422 | 0.151 |
| Logistic regression (all) | 0.741 (0.726–0.758) | 0.495 | 0.144 |
| Gradient boosting | 0.772 (0.758–0.787) | 0.548 | 0.137 |
Gradient boosting ranks accounts with AUC 0.772, against 0.672 for the rule and 0.741 for logistic regression. The gain over logistic regression is 0.031 (95% CI 0.021 to 0.041).
Where it misses
287 of the 1,327 test defaulters scored low. 92% of them had never paid late, used 28% of their limit (median) and had a median limit of 190,000. Six months of statements do not show their trouble coming.
Fairness and calibration
Data and method
- Source
- Yeh and Lien (2009), Default of Credit Card Clients, UCI ML Repository 350, CC BY 4.0
- Size
- 30,000 accounts, Taiwan, April to September 2005; 22.1% defaulted
- Split
- Stratified 80/20: 24,000 train, 6,000 test; 5-fold CV on train
- Audit
- 345 undocumented education codes, 17,496 undocumented payment-status values, kept and listed
Four methods: the late-last-month rule, logistic regression on last month's status only, logistic regression on all fields, and gradient boosting. Confidence intervals from 1,000 bootstrap resamples; the model comparison uses a paired bootstrap.
| Signal | AUC lost when shuffled |
|---|---|
| Repayment status last month | 0.026 |
| Months late of the last six | 0.015 |
| Worst delay in six months | 0.012 |
| Latest bill amount | 0.011 |
| Share of limit used | 0.010 |
| Payment two months ago | 0.006 |
| Credit limit | 0.006 |
| PAY_AMT4 | 0.003 |
Assumptions and limitations
- A missed payment costs five times a wasted call. The ranking holds at every ratio from 1:1 to 15:1
- Taiwan in 2005, after a card-debt crisis; an Australian book will have a lower base rate and needs refitting on its own history
- The data has no income, employment or bureau signals, which is why first-time defaulters are missed
- The effect of calling early is not measured. A trial that calls half the top lane is the way to measure it
What a lender would do next
- Weeks 1 to 2: export 24 months of statements, payments and arrears from the card platform
- Weeks 3 to 4: refit, check calibration and group rates, agree lane sizes with the collections lead
- Next two cycles: run the lanes with half of the top lane as a control
- Then: keep, resize or stop based on arrears roll rates
Technical depth
Features are built per account from the six-month window: months late, worst delay, bill and payment amounts and limit use. Thresholds are chosen on expected cost, not accuracy, which is why the call lane is larger than the base rate. Permutation importance is computed on the test set.