LLM-Based Corporate Actions Event Processing and Payment Validation System

Authors:

Hemachandru K

Department of Computational Technology SRM Institute of Science and Technology,

Kattankulathur, Tamil Nadu, India K.hemachandru@gmail.com

Saravanan Santhanam

Associate Professor

Department of Computational Technology SRM Institute of Science and Technology,)

Kattankulathur, Tamil Nadu, India saravans2@srmist.edu.in

Abstract—Financial institutions have a lot of operational work to meet when companies perform transactions such as paying dividends in cash, splitting stock, rights, merger and consolidation, and meeting downstream financial commitments in a correct and timely manner.Identifying record date, calculating shareholder entitlement, dividend payout as well as regulatory reporting during specified operating dates are all downstream requirements in which each announcement triggers. Manual processing at scale is not only impracticable, but also likely to contain errors, given the magnitude of corporate action notices that are generated on a trading session basis. These announcements consist of thousands of listed organisations on exchanges around the world. Meanwhile, all this is occurring at a time when banks and other financial organisations are still relying on archaic technology, i.e. human document reviewers and rule based software, which could be a liability during high demand times. The proposed architecture is a system which would employ an LLLM based pipeline to automate this operation. It utilizes a pre-trained language model to identify important financial data in announcements in any format (e.g. PDF, HTML, XML Feed, plan text file) and sorts the event into categories, and then calculates the payment liabilities associated with the event and finally writes all of the output data into a database after filtering against a set of domain application business rules. By comparing 1,200 announcements found in BSE and NSE, the system performed nearly 60 per cent less than the human-driven operations regarding accuracy (92.4 per cent of events outperforming) and throughput (320 documents/hour). The system is modular, consisting of Large Language Models, FinTech Automation, Payment Validation, Natural Language Processing, Financial Event Extraction, Transformer Models and Corporate Actions Processing, which can be updated individually when necessary.