Real-time mobile-money fraud detection
Block fraudulent transactions in milliseconds, without slowing down legitimate customers.
The challenge
Mobile money has become West Africa's financial backbone, with billions of transactions a year. That scale attracts sophisticated fraud: account takeover (SIM swap), social engineering, mule networks, coordinated cash-outs. Classic static rules catch known patterns but miss novel fraud and generate too many false positives.
Every unjustified alert blocks an honest customer; every missed fraud erodes trust and the bottom line. Control must happen in real time — before the transaction is approved — over massive volumes and fraud patterns that evolve continuously.
Our approach
Shift deploys a streaming anomaly-detection engine that scores every transaction in milliseconds. We combine a supervised model, trained on confirmed fraud history, with unsupervised methods that flag never-before-seen atypical behaviour, to cover emerging fraud.
The system's strength lies in graph features: by modelling the network of accounts, devices, numbers and beneficiaries, we detect mule rings, star-shaped collection patterns and abnormal money-circulation velocities — signals invisible at the level of a single transaction.
The system produces a risk score, an explainable reason and a recommended action (allow, challenge with step-up authentication, block), all inside a feedback loop where fraud analysts label cases, continuously retraining the models to keep up with evolving threats.
Architecture
- Ingestion: real-time event stream (Kafka), on-the-fly context enrichment
- Models: supervised gradient boosting + unsupervised Isolation Forest / autoencoder
- Graph: network features (GNN) over accounts, devices, beneficiaries
- Decision: <50 ms scoring, hybrid rules engine, analyst labelling loop
A mobile-money provider halves its fraud losses while cutting wrongful blocks in half, protecting both margin and customer experience.