Detecting Money Laundering in Rwandan Mobile Money: A Machine Learning Framework

Battle Count: 76

The paper addresses AML transaction monitoring in mobile money, not quantitative trading. However, there are tangential connections: the use of gradient boosting (LightGBM) for imbalanced classification, anomaly detection techniques, and the handling of extreme class imbalance (~0.1% prevalence) are methodologically relevant to fraud detection in trading systems. The operational alert-budget framework and PR-AUC evaluation under rare events could inform risk management in algorithmic trading. The causal feature engineering approach (rolling velocity, burstiness, counterparty diversity) shares conceptual parallels with market microstructure features. Overall relevance is low as the domain, objectives, and deployment context differ substantially from quantitative trading.

Moderate complexity. The pipeline involves: (1) causal feature engineering with multiple time windows (1h, 24h, 7d) requiring careful temporal aggregation without look-ahead leakage; (2) training and tuning of 6 base models (LR, RF, LightGBM, IF, LOF, Autoencoder); (3) logistic-regression late-fusion stacking; (4) threshold calibration at ~90% precision; (5) four-band risk scoring mapped to analyst workflows. However, all components use standard ML libraries (scikit-learn, LightGBM, PyTorch/Keras for autoencoder). The main complexity lies in the feature engineering under strict causal constraints and the operational calibration protocol rather than in novel algorithmic implementation. The synthetic data simplifies preprocessing (no missing values by construction).

SAML-D dataset is publicly available on Kaggle. Feature engineering, model training, and evaluation code are stated as available on request pending institutional review. Hyperparameters are specified (LightGBM: lr 0.02-0.05, num_leaves=64, min_data_in_leaf=200, feature_fraction=0.8; RF: 500 trees; Autoencoder: [128,32,8,32,128] with d=8 latent). Chronological 70/15/15 split is described. However, no GitHub repository is provided, single seed used, and no confidence intervals reported. The paper explicitly notes results are on synthetic data and not production-calibrated.