'Buy Now, Pay Later' (BNPL) Scoring at Checkout
A deferred-payment fintech decides in 200 ms, inside the checkout of 5,000 online shops, whether to approve splitting a purchase. 60% of users have no traditional banking history. Each point of improvement in default detection is worth €3M a year. Under European AI regulation, credit scoring is high-risk: reasons must be given for each denial.
Where the money comes from
What to do now
- 1Calculate margin per approved transaction and average loss per default by shop.
- 2Train the model only with variables available at checkout.
- 3Economic threshold per shop and denial reasons (SHAP).
For your committee
One page with the decision, the impact, the risks, the steps and the KPI to track.
The print dialog opens: choose "Save as PDF".
Watch out: more volume is not always more profit
Shops push to approve more; each extra approval point on marginal profiles raises defaults by more than the commission. The threshold is set by margin per transaction, different for each shop.
KPI for the committee
Default loss over volume and approval rate by shop.
Detail for your data team
Proposed technique: XGBoost
Why: It is a massive tabular problem, with non-linear relationships between many signals and where each point of accuracy is worth millions. XGBoost is the standard for this type of data, responds in milliseconds and, with SHAP and monotonic constraints, delivers reasons for each decision to comply with regulation.
Cost of getting it wrong: Approving a defaulter costs the amount financed; denying a good customer loses the sale and the shop's commission, and upsets the merchant.
Data needed: History of 30M transactions with payment outcome (paid / defaulted).; Transaction data: amount, shop, category, time, device.; User behaviour: previous purchases, returns, on-time payments.; Alternative signals: age of the email address, address consistency.