GA, UNITED STATES, November 20, 2025 /EINPresswire.com/ -- Researchers have developed a graph-based expert system that improves the accuracy and prediction stability of automated bank reconciliation. By modelling historical transactional data as a network graph, the system can learn complex one-to-many matching scenarios that existing tools often fail to predict correctly. The findings point to more reliable automation for high-risk domains such as finance and accounting.

Bank reconciliation is an essential part of maintaining the financial health of a business, requiring bookkeepers to match incoming bank statement lines to invoices. For large businesses that process thousands of records, it is both time-consuming and tedious, which is why many rely on automated tools that suggest likely matches for bookkeepers to confirm. While these tools work reasonably well for simple one-to-one matches, they often perform poorly when a single payment needs to be reconciled against multiple invoices (one-to-many matches).

In a new study published in The Journal of Finance and Data Science, a team of Australian researchers explored whether graph representation learning could improve the accuracy of match suggestions in these scenarios.

"Instead of modelling each transaction in isolation, a system could leverage a network mapping out the entire general ledger, where each historical record and its reconciliation are represented as a node and edge in a graph." shares Justin Munoz, lead author of the study. "New records can then be added to this graph, transformed into numerical representations or embeddings, and fed into a downstream machine learning model that scores the match likelihood for any pair of records."

Trained and evaluated on three years of real-world bookkeeping data, the graph-based method was shown to significantly improve match accuracy, outperforming an industry standard, with the largest gains on one-to-many matches. The researchers attributed these gains to higher-quality embeddings that capture both the structural properties of the ledger graph and the contextual information contained in transactions.

Further, the team found that graph-based models exhibited much lower prediction instability than other non-graph embedding methods such as Google's BERT, a popular language model. In this context, prediction instability refers to variation in model performance when a model is retrained multiple times. As shown in Figure 1, the best models cluster in the top-left region of high accuracy and low prediction instability.

"For high-risk domains such as finance and accounting, stability matters just as much as accuracy." adds Munoz. "Our findings highlight a promising direction for accounting technology that bookkeepers can rely on in day-to-day work, improving both trust and reliability."

Lucy Wang
BioDesign Research
email us here

Legal Disclaimer:

EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.

Information contained on this page is provided by an independent third-party content provider. Frankly and this Site make no warranties or representations in connection therewith. If you are affiliated with this page and would like it removed please contact [email protected]