Detecting unusual transactions in a mobile-money service

Context

Compliance checks relied on samples and fixed rules, unable to keep up with growing volumes or spot new patterns.

Our solution

• Preparation and historisation of transaction data
• Anomaly-detection model combining business rules and unsupervised learning
• An explainable score for every alert
• Prioritised review queue integrated into the analysts' tool

Results

• Suspicious cases detected earlier
• Review time halved
• False-positive rate tracked weekly

« Our analysts now spend their time on the real cases. »