Key takeaways
- Collusion fraud hides in the links between accounts, so a model that scores each account alone misses it.
- H2O Driverless AI combines link features, such as shared devices, addresses and counterparties, with your own expert features in one model.
- Every flagged group comes with reason codes your investigators and model risk team can read.
- The scoring layer runs beside your existing fraud engine, so your rules and case process stay in place.
Each account looks normal on its own
A colluding buyer and seller can each pass every check your rules apply. The fraud only shows when you see that they share a device, pay each other in a loop, or were opened from the same address in the same week.
Transaction volumes and thresholds vary by market and season, so fixed limits either fire too often or not at all, and your investigators spend their time joining the dots by hand. When you act on a ring, you close several accounts at once, and you need evidence you can show to a customer, a regulator and your own model risk team.