DATA-DRIVEN FRAUD PREVENTION
10 signals.
Explainable rules.
A detection system calibrated against percentile distributions of real data, rather than thresholds chosen by intuition.
Explore the technical approach
- Challenge
- Distinguish suspicious behavior from legitimate errors without overwhelming the team with alerts.
- Implementation
- Analysis of a sample of approximately 3,000 payments and 1,232 users. Percentile calibration, cross-provider decline code reviews and investigation of account connections.
- A decision that mattered
- One signal was removed because the data showed it could not separate legitimate from fraudulent accounts. The goal was evidence quality, not a larger rule count.
