Summary
This research explores how Large Language Models can transform the automated categorisation of SME bank transactions. Traditional rule-based systems struggle with the diverse and unstructured nature of SME transaction descriptions, leading to inconsistent classification that hampers effective financial analysis.
The study proposes a novel approach that leverages LLMs’ natural language understanding capabilities to improve categorisation accuracy and robustness. The methodology specifically addresses challenges unique to SME transaction data, including varied descriptions, industry-specific terminology, and irregular payment patterns.
Experimental evaluation on real SME transaction datasets shows significant improvements in categorisation accuracy compared to traditional approaches, with particular strength in handling ambiguous and previously unseen transaction types.
Authors: Brandi Jess (SME Capital, University of Warwick), Matt Arderne (sea.dev), Pietro Alessandro Aluffi (SME Capital, University of Warwick), Daniel Rodrigues (SME Capital), Marya Bazzi (sea.dev, University of Warwick, The Alan Turing Institute), Kate Kennedy (SME Capital), Martin Lotz (University of Warwick)