Resource-Efficient Anti-Money Laundering Detection: A Tabular Machine Learning Study for Compute-Constrained African Financial Systems

Money laundering costs an estimated $0.8–2.0 trillion annually, and the financial institutions most exposed to this risk in emerging markets are often the least able to afford commercial, compute-intensive detection systems. We present a systematic study of automated anti-money-laundering (AML) detection on the IBM AMLworld HI-Small benchmark: five classical machine learning algorithms, a multi-layer perceptron (MLP, 128→64→32), and three model-compression techniques (feature pruning 13→9, INT8 quantization 4×, k-centre coreset selection). On a 500k-row stratified sample (test set: 100k rows, 1,783 fraud cases, 1.78% prevalence), Random Forest achieves F1 = 0.3445 (AUC = 0.9330); XGBoost achieves F1 = 0.2756 (AUC = 0.9397); and an MLP with validation-set threshold tuning achieves F1 = 0.3553 (AUC = 0.9332). A combined pruning-plus-quantization pipeline reaches F1 = 0.3560 at a 75% reduction in model size (95.0 KB → 23.8 KB). We additionally quantify the cost of single-institution training: restricting training data to a single simulated bank collapses MLP F1 by 0.284, while Random Forest is comparatively robust (+0.019). We argue these findings speak directly to AML deployment realities in African financial systems, where cross-institutional and cross-border data sharing is fragmented and many payment service providers and mobile-money operators cannot afford GPU-scale compute or costly vendor compliance software.