surafelasfawosen/NBE-Ethiopia-AML-Detection

An AI-driven fraud detection system tailored for the National Bank of Ethiopia. Implements NID Entity Tracking, unsupervised anomaly detection, and cost-sensitive supervised learning to catch money mules in a 7-million row dataset with 0.05% extreme class imbalance. # NBE Anti-Money Laundering (AML) Deep AI Engine **Empowering the National Bank of Ethiopia and the Ethiopian Financial Ecosystem with Next-Generation AI Fraud Detection.** Multi-layered Anti-Money Laundering (AML) detection pipeline tailored to the National Bank of Ethiopia (NBE) compliance guidelines. This system is designed to intercept complex money laundering typologies, including structuring (smurfing) and sophisticated pass-through money mule networks, despite extreme data class imbalances. --- ## Dataset & Ethiopian Localization Instead of using the raw Kaggle IBM AML dataset as-is, the data was heavily engineered and localized to accurately reflect the Ethiopian banking ecosystem. All transactions were mathematically scaled to realistic Ethiopian Birr (ETB) distributions and mapped to local institutions: * **Total Transactions:** `6,924,049` records * **Unique Customer Profiles:** `423,543` synthesized Ethiopian National IDs (NIDs), complete with localized metadata (e.g., Ethiopian names, customer segments). * **Class Imbalance (Laundering Ratio):** `~0.05%` (Only 3,565 positive laundering cases out of nearly 7 million legitimate transfers) * **Market Scale:** Transactions are dynamically mapped across 16 commercial Ethiopian banks, with the Commercial Bank of Ethiopia (CBE) holding a realistic ~66.0% market share, followed by Awash, Dashen, and others. --- ## Core Innovation: Dual-Dataset Architecture (Transactions + KYC) A major flaw in traditional AML models is their reliance on a single, isolated transaction dataset. To solve this, we architected a **Dual-Dataset System** by generating and merging a dedicated **KYC (Know Your Customer) / NID Profile Library**. By cross-referencing raw transactions against this secondary KYC dataset, the model stops looking at isolated accounts and starts tracking **Personal Entities**. We map every bank account to a unique 12-digit Ethiopian National ID, unlocking powerful historical tracking capab …