CASE STUDY — CREDIT UNDERWRITING
A better model, a simpler pipeline.
For repeat cash-advance customers, a small-business lender needed to know who was likely to default on their next draw. NEXUS out-predicted the incumbent model — with a fraction of the engineering behind it.
The Task
Repeat-draw default risk
The Pipeline
One page, not thousands of columns
A model too complex to trust
Once a small business has drawn a cash advance and started repaying it, every new draw is its own credit decision — and getting it wrong compounds fast across a repeat-borrower book. The lender’s incumbent model chased accuracy with a feature pipeline running into the thousands of columns: powerful in principle, but a black box in practice. No credit officer could look at it and say why an applicant was flagged. The lender needed a model that was both more accurate and something a human could actually reason about.
Fewer features, more understanding
NEXUS was given the same underlying signals as the incumbent — bank activity, credit history and each customer’s past repayment behavior — but instead of engineering thousands of narrow features to compensate for a model with no real understanding of lending, NEXUS brought that understanding with it. The result is a compact, human-readable feature set that a credit officer can review on a single page, with none of the accuracy traded away to get there.
Bank activity
Cash flow patterns from the customer's connected accounts
Credit history
Standard bureau data, the same inputs the incumbent used
Repayment behavior
How this customer has actually repaid past draws
03 — THE RESULT
more accurate risk ranking than the production model
Measured on the client’s own evaluation, with a fraction of the incumbent’s feature pipeline — and a model a credit officer can actually read.
Production model
Thousands of features
NEXUS
One page, +18% accuracy
04 — THE IMPACT
Accuracy and clarity, not a trade-off.
Lenders have long assumed that a more accurate model means a harder one to explain. NEXUS breaks that trade-off: better risk decisions, built on a pipeline simple enough for a credit officer to review line by line. That’s a model risk teams can approve, examiners can understand, and the business can actually trust in production.
