Spaces:
Runtime error
Apply for community grant: Academic project (gpu and storage)
Leveraging Large Language Models for SWIFT MT564 Corporate Action Anomaly Detection⋆
Author :Paresh Mishra1, Guide -Sudha BG2
1 Department of Computer Science, PES University, Electronic City, Bangalore, 560100, Karnataka, India,
paresh.mishra23@gmail.com : sudha.bg@greatlearning.in, WWW home page: https://www.pes.edu
2 Department of Computer Science, PES University, Electronic City, Bangalore, 560100, Karnataka, India
Abstract. This study aims to harness the advanced capabilities of large language models to automate the Anomalies detection of Swift Messages starting with Corporate action messages as MT564 with ISO 15022 For- mat The SWIFT network is integral to global financial transactions, routing millions of messages each day—among them MT564 Corporate Action Notifications, which inform custodians and investors of criti- cal events such as dividends, mergers, and redemptions. Anomalies in these messages, resulting from fraud, syntax errors, or compliance lapses, threaten both financial integrity and regulatory compliance. Traditional rule-based and machine learning fraud detection systems optimized for tabular data struggle with the SWIFT network’s semi-structured, context- rich format and rarity of labeled anomalies. This work presents a novel, reproducible anomaly detection pipeline, combining a deterministic SWIFT parser, a fine-tuned mt564-gemma-lora model with Low-Rank Adapta- tion (LoRA), and a lightweight AI agent for real-time anomaly scoring and alerting. Evaluated against a robust holdout set and in production- like shadow deployments, mt564-gemma-lora demonstrates state-of-the- art accuracy and scalability, providing a framework for dependable cor- porate action surveillance in modern financial operations.
Keywords: SWIFT, MT564, Anomaly Detection, LLM, LoRA, Finan- cial Messaging