LLM-STARS: LLM-enhanced standardization of time-series analysis and relationships in subledgers
2025
Financial accounting systems rely heavily on subledgers to track detailed transaction records. However, modern systems often evolve into complex architectures where different components use inconsistent labeling conventions, making it difficult to understand and utilize important relationships within subledger data. This paper presents a novel framework LLM-STARS (LLM-Enhanced Standardization of Time-series Analysis and Relationships in Subledgers) that leverages Large Language Models to enhance time series analysis of subledger data through relationship modeling. LLM-STARS represents subledger data as a graph where financial events connect accounting segments, while firstly using LLMs to generate standardized interpretations of these events based on both their attributes and their role in moving money through the accounting system and then explicitly modeling relationships among subledger time series. The framework effectively identifies two types of relationships between subledger activities: reconciliation relationships that capture clearing/settlement patterns, and similar pattern relationships that reflect shared business drivers. We demonstrate through extensive experiments on enterprise testing data representative of real-world usage patterns that incorporating these relationships significantly improves the subledger data analysis performance as compared to traditional univariate approaches. For example, LLM-STARS improves anomaly detection F1 score from 0.516 to 0.621 (by 20.3%) for collective seasonal outliers and decreases symmetric mean absolute percentage error from 53.82 to 27.83 (by 48.3%). Moreover, LLM-STARS provides interpretable results through with language descriptions while maintaining the technical rigor required for financial applications.
Research areas