Manaz Afzal
Banking and lending · Collections

A follow-up queue that calls before the payment is missed

A monthly scoring step that sorts every card account into call, reminder or routine lanes. Tested on 6,000 accounts it never saw: the riskiest fifth holds half of next month's missed payments.

50%of next month’s missed payments in the top 20%
0.772AUC, against 0.672 for the late-last-month rule
17%lower modelled cost than the rule at 5:1
92%of missed defaulters had never paid late

Executive summary

A monthly score puts half of next month’s missed card payments into the riskiest fifth of accounts, so a two-person collections team calls before the payment is missed rather than after.

50%of next month’s missed payments in the top 20% of 6,000 held-out accounts
0.772AUC (95% CI 0.758 to 0.787) against 0.672 for the late-last-month rule
17%lower modelled cost than the rule at a 5:1 miss-to-call cost

Situation

Collections teams call anyone late last month. By then hardship options are fewer and the account is already in arrears reporting. The team cannot call every account.

Approach

Gradient boosting on six months of bills, payments and late flags for 30,000 card accounts, tested on 6,000 held out, with call, reminder and routine lanes cut on expected cost.

Result

The call lane goes on to miss at 55%, the routine lane at 9%. Ranking quality is similar across groups (AUC 0.75 to 0.79). First-time defaulters remain hard: 92% of missed defaulters had never paid late.

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.

Statements closeSix months of bills,payments, late flagsScoreRisk of missing nextpayment, with reasonsLanesCall, reminderor routineCRM and diallerCall list and SMSreminders queuedOutcomePaid, plan agreed ormissed; logged
Monthly workflow. A person approves the call list; the score never changes a limit or declines credit.
LaneAccountsWent on to missShare of all misses
Call this week (top 20%)1,20055%50%
Reminder (next 30%)1,80022%29%
Routine (bottom 50%)3,0009%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

0 months late19,931 accounts12%1 month late4,426 accounts30%2 months late1,899 accounts39%3 months late1,154 accounts51%4 months late951 accounts57%5 months late298 accounts57%6 months late1,341 accounts70%All accounts 22.1%
Two in three of accounts late in all six months defaulted. Most accounts (19,931) were never late, and they still produce a third of the defaults. Data: UCI Default of Credit Card Clients, 30,000 accounts
Score tenth 1600 accounts69%Score tenth 2600 accounts40%Score tenth 3600 accounts24%Score tenth 4600 accounts20%Score tenth 5600 accounts20%Score tenth 6600 accounts14%Score tenth 7600 accounts10%Score tenth 8600 accounts12%Score tenth 9600 accounts7%Score tenth 10600 accounts4%Base rate
The top score tenth defaulted at 69%, the bottom at 3.5%. The riskiest 20% holds 50% of next month's missed payments.

Cheaper than the rule at the same team size

3,0004,0005,0006,0007,0000.00.20.40.60.8Score, by cut-offLate-last-month ruleBest cut-off 0.15:cost 3,353 against 4,062Score cut-off for a call (share of risk)
Cost counts a missed payment as five times a wasted call. The best cut-off lowers modelled cost by 17% against the late-last-month rule; at ratios from 1:1 to 15:1 the score is 8% to 60% cheaper.
MethodAUC (95% CI)PR-AUCBrier
Rule: late last month0.672 (0.658–0.687)0.3510.228
Logistic (PAY_1 only)0.676 (0.659–0.695)0.4220.151
Logistic regression (all)0.741 (0.726–0.758)0.4950.144
Gradient boosting0.772 (0.758–0.787)0.5480.137

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.

≤02,598 accounts25%0–10%8,644 accounts17%10–30%3,551 accounts17%30–50%2,833 accounts23%50–70%2,852 accounts25%70–90%3,474 accounts28%90–100%3,933 accounts24%>100%2,115 accounts30%
Limit use matters less than payment history: default rises from 17% to 30% across the range.

Fairness and calibration

Base rateShare flaggedSex: Male53%Sex: Female47%Age: [20, 25)60%Age: [25, 30)47%Age: [30, 35)44%Age: [35, 40)47%Age: [40, 45)48%Age: [45, 50)56%Age: [50, 60)59%Education: Graduate school42%Education: University53%Education: High school57%
Flag rates at the cost-optimal cut-off (which flags 49% of accounts) range from 42% to 60% across groups; the youngest and oldest age bands and high-school-educated customers are flagged most. Ranking quality is similar in every group (AUC 0.75 to 0.79). A person reviews these rates monthly.
0%20%40%60%0%20%40%60%Perfect calibrationGradient boostingPredicted default rate (tenth average)
Predicted and observed rates stay within four points in every tenth (Brier 0.137), so a cut-off can be set from a cost ratio.

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.

SignalAUC lost when shuffled
Repayment status last month0.026
Months late of the last six0.015
Worst delay in six months0.012
Latest bill amount0.011
Share of limit used0.010
Payment two months ago0.006
Credit limit0.006
PAY_AMT40.003

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.

Financial impact

About 1,000 accounts a month kept out of arrears on a 200,000-account book, if early calls work as assumed

The effect of calling early is not measured in the data. The figure below shows what the ranking is worth under a stated conversion assumption, to be tested with the control group in the rollout plan.

InputValueSource
Card or loan accounts200,000 bank or consumer lender
Accounts missing a payment each month5% an Australian book sits well below the 22% in this data
Misses in the top 20% call lane50%Measured
Reached accounts that avoid arrears after an early call20% to be measured in the trial
Collections cost avoided per account kept currentA$150 calls, letters, hardship handling

Calculation

  1. Misses a month10,000
  2. Reached before the due date5,000
  3. Kept out of arrears1,000
  4. Annual collections cost avoidedA$1,800,000

Annual value, pending the control-group trialA$1,800,000

Lower credit losses and better customer outcomes under hardship obligations are not counted.

Implementation and rollout

Two statement cycles with half the call lane held as a control

  1. Weeks 1 to 2

    Export

    24 months of statements, payments and arrears from the card platform.

    Collections lead, Manaz

  2. Weeks 3 to 4

    Refit and review

    Refit, check calibration and flag rates by group, agree lane sizes with the collections lead.

    Manaz, risk and compliance

  3. Cycles 1 to 2

    Controlled run

    Run the lanes with half the top lane as a control; reasons written to the CRM call list.

    Collections team

  4. Then

    Decide

    Keep, resize or stop based on arrears roll rates.

    Head of credit

Tools

Core banking or card platform export, Salesforce or HubSpot call lists, Dialler, Python and scikit-learn monthly job

Team

  • Head of credit owns lane sizes and the decision
  • Collections lead reviews the list before calls start
  • Risk and compliance monthly group-rate review
  • Manaz model, CRM integration, trial analysis

The timeline is a plan for a typical engagement; it depends on data access and how many systems are involved.

Risks and mitigations

What could go wrong in production, and the control for each

RiskMitigation
Score used to change limits or decline creditScores order calls only; written policy and a person approves the list.
Some groups flagged more oftenFlag rates by group reviewed monthly (42% to 60% at the cost-optimal cut-off).
Taiwan 2005 data does not transferRefit on the lender’s own 24 months; base rate and signals differ.
First-time defaulters missedAdd income or bureau signals where permitted; routine lane still receives standard reminders.

Key takeaways

Key takeaways

  • Payment history in the last six months carries most of the signal; limit use matters less.
  • Cut lanes on expected cost, not accuracy; the call lane is larger than the base rate for that reason.
  • Most accounts were never late, yet they produce a third of defaults, which no history-based score can see.
  • The value of early calls must be measured with a control group before scaling.