Adaptive Graph-Based Risk Scoring for Real-Time Instant Payment Systems

Instant payment rails and mobile money platforms such as UPI, Pix, and FedNow require fraud decisions in a few hundred milliseconds, but most deployed and published fraud models are batch oriented and sequence only. They are trained and evaluated offline, treat each account in isolation, and provide limited visibility into mule networks and collusive cash out structures. This paper presents a graph based streaming risk scoring architecture that brings graph neural network (GNN) style inference into the authorization path while respecting a sub 300 ms end to end service level objective. The system maintains a dynamic transaction graph over accounts, devices, merchants, and institutions, updates this graph as payments arrive, and uses an online GNN to propagate risk so that emerging mule rings can be flagged even when individual transfers appear benign. Raw scores are calibrated into three operational actions, allow, step up, and block, under explicit fraud loss and customer friction budgets. Using a labeled synthetic mobile money dataset representative of instant payment rails, we build a real time risk engine around this architecture and quantify its latency, queueing behaviour, and decision quality. A strong tabular baseline representative of existing non graph deployments is implemented inside the same engine and evaluated side by side, providing a concrete comparative analysis between current practice and the proposed architecture.