Enhancing Financial Crime Identification in Traditional Banking through Semi-Supervised Anomaly Detection

Financial crimes pose significant risks to financial institutions, national economies, and society as a whole. The major threats in this domain are addressed through specific regulatory frameworks, such as Anti Money Laundering (AML), Combating the Financing of Terrorism (CFT) and opposing the proliferation of weapons of mass destruction. To mitigate these risks, supervisory authorities impose substantial penalties on institutions that do not adopt effective measures of prevention. Consequently, banks are required to comply with strict regulations designed to detect and prevent such crimes. Although these tasks have traditionally been handled by rule-based Transaction Monitoring (TxM) systems, recent advances in Machine Learning have introduced new paradigms to address this complex challenge. This thesis presents a real-world case study on the adoption of unsupervised and semi-supervised Machine Learning techniques for anomaly detection in the context of a traditional banking system, a sector typically characterized by strong conservatism. In this work, the "Multicriteria Anomaly Detection" (MAD) project is described in detail: using approximately 3.3 billion anonymized transactions collected over 12 months by an Italian financial institution, a Machine Learning pipeline for TxM was developed, optimized, and deployed in production. The work spans the entire process: from the presentation of the datasets, to the extraction of 98 aggregated features describing each account’s monthly transactional behavior, to the training and hyperparameter tuning of a semi-supervised AutoEncoder, and finally to the evaluation of its performance with dedicated metrics. Comparative analyses were also carried out against other unsupervised anomaly detection models, specifically One-Class Support Vector Machine, Isolation Forest, and Local Outlier Factor. This exploration is particularly relevant given the increasing demand for more effective approaches to counter financial crimes and highlights how the application of Machine Learning can contribute to the evolution of monitoring systems in this critical sector.