Machine Learning Framework Developed to Combat Mobile Money Laundering in Rwanda

This research paper introduces a sophisticated machine learning framework designed to detect money laundering and terrorism financing within Rwanda's mobile money ecosystem. Given the widespread adoption of mobile money across Sub-Saharan Africa, these platforms, characterized by high volumes of low-value transactions, have unfortunately become susceptible to illicit financial activities. Rwanda serves as a pertinent example, with millions of active users on networks like MTN and Airtel, presenting a challenge for traditional rule-based monitoring by its Financial Intelligence Centre (FIC).

The framework addresses specific challenges inherent in the Rwandan context, including extreme class imbalance (where illicit transactions are very rare), a scarcity of timely labeled data, and limited investigator capacity. It utilizes a synthetic dataset, SAML-D, containing nearly 10 million transactions with 17 money laundering typologies. The researchers engineered various account-centric behavioral features, such as transaction velocity, net-flow directionality, and counterparty diversity, to effectively identify suspicious patterns.

The study benchmarks several supervised and unsupervised machine learning models, including Logistic Regression, Random Forest, LightGBM, Isolation Forest, and a dense autoencoder, along with a late-fusion meta-learner. The evaluation focuses on operational metrics relevant to financial intelligence units, such as PR-AUC, recall at high precision, and alerts per 10,000 transactions. The LightGBM model and the fusion stacker demonstrated strong performance, effectively identifying a significant number of laundering cases with high precision and manageable alert rates.

The paper emphasizes the operational contribution of the framework: a governance-aware pipeline and evaluation protocol specifically tailored to the constraints faced by an African mobile money regulator. It outlines a practical path for transitioning from synthetic data prototyping to real-data validation in collaboration with institutions like the National Bank of Rwanda and the FIC, offering a concrete tool for enhancing financial security in the region.