Business case, 28 September 2026
Managing risk on 6/9/12
Edvin Maehre
The main risk question
How do we let customers buy bigger baskets over longer terms, when each bad approval costs more and takes longer to show?
Answer: customer-level data improves risk ranking; a control group proves the NPV.
Add customer data to the score
Bureau data and Scalapay history rank risk better than the internal risk score alone: Gini 35 to 45.
Slide 6
Read each cohort in its first months
The first missed instalment shows the direction long before a 12-month loan matures.
Slides 4 and 7
Prove the NPV on a control group
Today's cutoff is already tight. A control group measures what approving more customers earns.
Slide 7
2
Risk grows Scalapay's total profit
6/9/12: the first product built on the customer view
Customer-level data sharpens each decision at checkout. A control group shows where approving more pays, and whether the term itself adds risk. We read each cohort in its first months.
Scalability
Every product and market reuses one view of each customer, so each launch starts from what we already know.
Performance
Every rule, data source and provider keeps proving its worth, or we remove it.
One view of each customer
Assumption to verify: one credit limit per customer, shared by all products
Judge each decision and each customer by their expected NPV.
3
The score ranks risk overall, but not in order in the middle
Missed rate by score decile (10 = lowest risk)
8.1% vs 3.3%
Decile 4 misses 2.4 times as often as the safer-looking decile 5.
7.9% vs 3.4%
Decile 6 misses 2.3 times as often as decile 7.
The book is too young to show losses: 83% of orders are under two months old, and 72 are in default. So we analyse loans through the first missed instalment, at the same loan age for every plan.
Something the score may not fully incorporate drives this risk (slide 5).
Missed: first instalment after checkout overdue or paid 15+ days late. 5,385 orders with a score reached it by 23 Sep 2026.
4
The opportunity: incorporating more customer-level data
Bureau: depth of credit history
4.6% vs 1.2%
Same score (decile 9): customers the bureau does not find miss 4 times as often as customers it finds.
The number of credit lines alone ranks risk almost as well as the score.
Scalapay: money still owed
9.3% vs 2.7%
Orders placed while the customer still owes on another Scalapay plan miss 3.4 times as often. They miss more in every score band.
They explain the break at decile 4 (slide 4): without them, decile 4 misses 3.7%, like decile 5.
The score places these orders only a little lower.
Both describe the customer, not the order: the base of one view of each customer.
Missed: first instalment after checkout overdue or paid 15+ days late. Bureau: decile 9, 1,199 orders. Open plans: 389 of 5,385 scored orders still owed on another plan (36 missed).
5
Customer data ranks risk better than the score alone
Proof of concept: a scorecard built on top of today's score, adding bureau data and Scalapay history.
Ranking power (Gini), on customers the model never saw
Scorecard: +9 points (95% range +3 to +16). Holds out of time: 26 to 45.
57 vs 48
misses caught of 172 at a 10% decline rate: the scorecard catches 19% more.
9.1% vs 5.5%
Orders only the scorecard declines miss 1.6 times as often as orders only the score declines.
Today's cutoff is already tight: the value is in approving more customers safely.
Declining more loses NPV for every model tested on this data. A control group beyond the risk frontier can measure what approving more earns.
Proof of concept: points scorecard, 5,385 orders with a score, 172 missed. 5-fold, each customer in one fold.
6
Grow 6/9/12 with a learning budget and stop rules
How each rollout learns
A learning budget
Mature products fund the learning on 6/9/12. Risk and the business agree its size.
A control group
Tests the scorecard, and whether the term adds risk. Also tests new data for thin files.
12-month orders miss 1.6 times as often at the same score; the test shows whether the plan or the customer causes it.
Stop rules
Each rollout has a warning level (mitigate) and a limit level (pause). The risk appetite sets both.
AI: an LLM does the first analysis of weak rules. An analyst then keeps, changes or removes each rule.
7
Roadmap for 6/9/12: the first four months
Compliance: the creditworthiness check is a legal duty; CCD II applies from 20 November 2026.
8
Where 6/9/12 leads: one view of each customer drives every decision
The customer
Sees their credit limit in real time, and how to raise it.
6/9/12: the first product built on it
Customer-level data sharpens every decision. A control group shows where to approve more.
Scalability
Every product, market and channel plugs into the same base.
Performance
Every rule and data source keeps proving its worth, or we remove it.
One view of each customer
One risk score per customer, updated on every significant event.
9
The customer side: what one view makes possible
Customers understand their credit
The customer sees their credit limit in real time, and what shapes it.
A reason to share data
The app shows how to raise the limit, for example by sharing bank data, and what sharing is worth.
An explanation on demand
We tell the customer everything we can within our regulatory duty: their situation and how to improve it.
First step: test open banking data on a subset of our customers. If it performs, we productise and incentivise sharing data directly in the app.
10
Appendix: One milestone, the same for every plan
Sample and outcome
18,321
orders in the file, to 23 Sep
5,507
reached instalment 2, not refunded
5,385
of them have a score
172
missed (3.2%)
Missed: instalment 2 overdue, or paid 15+ days late. One row = one order.
Left out, and why
Flag 34
Matches the 72 defaults exactly, so it is set after the outcome.
Status and repayment columns
They describe the outcome itself.
Demographics
Not used as inputs to the scorecard.
Data we lack
Declines, credit limits, MDR, cost of funds, and mature losses.
Missed any instalment, by month
Month
6
9
12
1
2.1%
2.1%
3.8%
2
3.7%
3.3%
6.5%
3
7.3%
2.6%
5.3%
4
6.1%
n<30
5.7%
After month 2, only the April to June cohorts remain.
Data cutoff 23 Sep 2026. Month = months after checkout, all cohorts pooled. Cells under 30 orders hidden.
11
Appendix: Same score decile, different risk
Bureau
Scalapay plan
Decile
Orders
Missed
Found
Not found
Still owes
Owes none
1
1%
17.4%
10.0%
n<30
n<30
11.9%
2
2%
7.4%
5.6%
10.8%
n<30
7.5%
3
3%
4.9%
4.5%
6.1%
n<30
5.1%
4
4%
8.1%
5.8%
13.6%
33.3%
3.7%
5
6%
3.3%
2.9%
4.5%
n<30
3.5%
6
9%
7.9%
7.5%
9.4%
23.4%
6.2%
7
14%
3.4%
2.8%
5.7%
2.7%
3.5%
8
18%
2.4%
2.0%
5.1%
5.5%
2.2%
9
22%
1.7%
1.2%
4.6%
4.8%
1.4%
10
21%
1.3%
1.2%
1.8%
1.9%
1.2%
One field alone (Gini)
Score decile
38.3
Total credit lines
33.9
Same 4,779 orders the bureau found.
Rating R1-R9: missed
R1 0.6%
R2 4.4%
R3 1.6%
R4 1.5%
R5 10.7%
R6 2.1%
R7 3.5%
R8 20.3%
R9 1.1%
Not in order. Rank correlation with the decile: 0.2.
Missed rates: 5,385 scored orders at the milestone, 172 missed. Orders: share of all 17,922 scored orders. Cells under 30 hidden.
12
Appendix: The gain holds on new customers and later orders
How it was built
Inputs
Score decile, plan, 7 bureau and 2 Scalapay history fields.
Bands
Each field cut into quartiles, plus a band for "missing".
Model
Logistic regression, shown as a points table.
Test
Split by customer: 5 folds, 5 seeds.
Ranking power (Gini)
Model
Mean (range)
Out of time
Score alone
35.4 (33.3–37.4)
25.5
+ Scalapay history
37.7 (35.4–39.5)
31.6
+ Bureau
44.0 (42.2–46.2)
42.2
+ Both (scorecard)
44.7 (42.7–46.9)
45.2
Rule set (tree)
33.2 (30.7–35.5)
41.7
Scorecard minus score: +9.4 (95% range +3.3 to +15.9, bootstrap by customer).
Top 5 inputs, spread of points
Grey: today's score. Black: bureau. Purple: Scalapay history. 20 points = the odds of paying on time double.
Most of the lift is bureau data; history adds a little.
5,385 scored orders, 172 missed, 4,995 customers. Out of time: tested on 1,347 orders from 2 Aug (44 missed).
13
Appendix: The ranking gain is measured; the money is not
Swap set at a 10% decline rate
Where the models disagree, the scorecard declines orders that miss 1.6 times as often.
NPV of declining 10% more
Model
NPV change
Score alone
EUR −18,504
Scorecard
EUR −13,445
Scorecard better by
EUR +5,059
Declining more loses money for every model: the best decline rate is 0–2%. A declined order needs about a 24% chance of an early miss before the decline pays.
Assumed: MDR 4%, 40% of misses default. With MDR 2–6% and 20–60% default, the scorecard is better by EUR +2,165 to +7,953.
The data holds only approved orders
A negative bureau event looks safe: 0.5% missed (2 of 424) against 3.4%. Policy picked these orders; we cannot see who it declined.
What measures the money
A control group beyond today's cutoff shows what approving more customers earns.
NPV on 5,385 scored orders, EUR 4.17m lent, against approving everything. The full case file holds EUR 10.9m.
14