Fraud Risk Propagation Through Account Networks¶
Industry: Financial Services | Role: Head of Financial Crime, Chief Risk Officer | Time to value: 3 hours
The Problem¶
When an account is confirmed fraudulent, every account that transacted with it becomes a potential risk. But fraud networks are not one hop deep — money moves through chains of intermediary accounts before reaching its destination. Identifying the full exposure from a single flagged account requires traversing a transaction graph that batch systems were not designed to explore.
The Traditional Approach¶
Transaction monitoring systems score accounts independently. When fraud is confirmed, an analyst manually traces recent transfers, flags direct counterparties, and opens cases. Network analysis, if it exists, runs as a separate batch job on a different platform. A typical investigation covers 2-3 hops and takes 4-8 analyst-hours. Accounts beyond the analyst's traversal window remain unscored. Meanwhile, clean accounts that happen to share a demographic with flagged ones get caught in broad-brush freezes.
With Uni¶
The notebook defines backward risk propagation through transfer edges: risk flows from flagged accounts to their counterparties, decaying with each hop. Accounts with no path to any flagged entity are explicitly isolated as clean using negation — not just unscored, but provably uninvolved. Every risk score carries a derivation trace showing the exact chain of transfers that produced it. The entire model is 12 declarative rules covering propagation, decay, clean-account isolation, and threshold classification.
What You'll See¶
- Risk-scored accounts across the full transaction network, not just direct counterparties
- A clean account whitelist — accounts provably isolated from flagged entities, reducing false freezes
- A propagation audit trail for each risk score, showing the exact transfer chain and decay calculation
Why It Matters¶
False positives cost financial institutions an estimated $50 per alert in analyst time. Replacing manual network tracing with automated propagation and provable clean-account isolation reduces both missed fraud and wasted investigation hours.