58%
Fewer false positives
23%
More fraud caught
<50ms
Decision latency
$24M
Annual loss reduction
The Challenge
What needed solving
The issuer's rules-based fraud engine was generating excessive false positives, frustrating customers and overwhelming the dispute team. At the same time, sophisticated fraud rings were evolving faster than the rules could be updated. Card decline rates had reached a level threatening customer retention.
Our Approach
How we partnered
We designed a real-time fraud scoring platform built around gradient-boosted models and graph-based features, deployed alongside the existing rules engine. Our data science team worked with the issuer's fraud analysts to engineer features from transaction velocity, merchant graph patterns, and device fingerprinting. We instrumented every decision so analysts could trace why a transaction was flagged.
Business Outcomes
The impact delivered
Within six months of go-live, the platform was making over 4 million real-time decisions per day with sub-50ms latency. Customer-impacting declines dropped sharply, NPS for the cards business improved by 11 points, and net fraud losses fell by $24M annually.
- 58% — Fewer false positives
- 23% — More fraud caught
- <50ms — Decision latency
- $24M — Annual loss reduction