Integration of Benchmark Analysis and Machine Learning for the Optimization of Sanctions Screening Testing

Sanctions screening is an important compliance process used to prevent financial crimes. This process relies on name screening algorithms, which must account for the fact that a name may appear as an alias or as another name variation. One of the central challenges in sanctions screening is the handling of weak aliases, which can increase the number of false negatives. This issue was identified within a business environment, where the screening process required data-driven support to address it more effectively. The aim of this thesis is to analyze sanctioned names and their aliases in order to develop a machine learning classifier for weak aliases, and to design a prototype data-driven application for managing and visualizing project data. The methodology includes a comparison of machine learning classification models, the evaluation of these models using different performance metrics, and the assessment of the application prototype based on business and technical requirements. In addition, the impact of the model on historical business data was examined in order to determine whether it had any effect on the recall performance of sanctions screening. The main results show that the selected model achieved balanced performance with a high F1-score. However, when applied to historical data, it led to a decrease in overall recall. Therefore, the model developed in this thesis contributes to supporting decision-making rather than serving as a system on which full reliance can be placed. The application, in turn, has the potential to improve project data management and visualization once it is applied within business processes and fully deployed.