To Wesley Leeroy, artificial intelligence can serve as a protector for the integrity of global finance.
As a Benjamin Franklin Scholar and recipient of the Amazon Future Engineering, Leeroy sought to demonstrate how AI can make a real-world impact through the presentation of his paper “AI-Enhanced Data Mining of the Financial Domain: Methods and Applications — Leveraging Deep Learning and NLP to Detect Anomalies and Combat Financial Crime,” at this fall’s IEEE International Conference on Data Mining (ICDM) in Washington, D.C. The paper was also published in the conference’s AI for Financial Crime Fight (AI4FCF) workshop.
“Penn Engineering’s interdisciplinary, collaborative, and ambitious environment made this possible,” says Leeroy. “From the start, I was encouraged to pursue a question that sits at the intersection of fields: How can we use AI not only to power new technologies, but also to help protect the integrity of global financial systems?”
Leeroy’s research tackles the challenge of detecting sophisticated financial crimes, often hidden in plain sight within the language of corporate disclosures. His framework integrates multiple deep learning architectures including graph neural networks (GNNs), convolutional neural networks (CNNs), and gated recurrent units (GRUs), to process and analyze vast amounts of unstructured financial text.
According to Leeoroy, the model was trained on regulatory filings from South American firms cross-listed in the U.S. Securities and Exchange Commission, analyzing thousands of filings from 2013 to 2023. By combining these textual insights with structured financial data, the model achieved an impressive accuracy rate of over 90% in its initial classification stage, Leeroy explains. Consequently, it identifies anomalous or inconsistent reporting behaviors that may signal fraud, misrepresentation, or regulatory evasion.
At the IEEE Conference, Leeroy presented the problem, solution, and impact. Regarding the problem, “legacy systems struggle to interpret the qualitative story of a company, overlooking contextual clues hidden in years of regulatory text.” As Leeroy’s research demonstrates, the solution rests in the “hybrid AI model that uses CNNs to detect local linguistic patterns, GRUs to track narrative evolution over time, and GNNs built with Google DeepMind’s graph neural network framework to understand relationships between entities.” As for the impact, “it’s a step toward proactive, intelligent financial surveillance, moving beyond static rule-based systems to dynamic, learning-driven detection.”