Anti-Money Laundering Compliance Using Feature Engineering with SQL Analytics, TF-IDF and Oversampling: Conditional Tabular Generative Adversarial Networks

Traditional Anti-Money Laundering (AML) systems rely on rule-based approaches, which often fail to adapt to evolving money laundering tactics and produce high false-positive rates, overwhelming compliance teams. This study proposes an innovative machine learning (ML) framework that leverages Conditional Tabular Generative Adversarial Networks (CTGANs) to address severe class imbalance, a common challenge in Suspicious Activity Reporting (SAR). Implemented in Python, CTGAN generates realistic synthetic samples to enhance minority-class representation, improving recall and F1-scores. For instance, the Random Forest (RF) model achieves a recall of 0.991 and an F1-score of 0.528 in oversampled datasets with engineered variables, highlighting the effectiveness of CTGAN in mitigating imbalance. This framework also incorporates SQL-based feature engineering using Oracle Analytics, creating dynamic variables such as cumulative sums, rolling averages, and ranks. The modelling phase and exploratory data analysis are conducted in the SAS programming language, employing Logistic Regression (LR) as baseline, Decision Trees (DT), and RF. Evaluation across undersampled and oversampled datasets, combined with varying probability thresholds, reveals key trade-offs between sensitivity and precision. Among the models, RF consistently achieves the highest ROC-AUC scores, ranging from 0.945 in undersampled datasets to 0.951 in oversampled configurations, demonstrating its robustness and accuracy in SAR detection. By integrating CTGAN and TF-IDF (textual feature transformation in Python) with SQL-engineered variables, this framework provides a comprehensive data-driven approach to AML. It reduces false positives, strengthens the detection of suspicious activities, and ensures scalability, adaptability, and compliance with regulatory standards.