| Abstract [eng] |
In the daily practice of corporate accounting departments, precisely and rapidly matching purchase and sales invoices stored in enterprise resource planning (ERP) systems with bank payment records remains a significant challenge. The heterogeneity of bank statement formats - driven by varied legal and technical requirements across jurisdictions - prevents full automation via rule‐based methods, since each client demands custom rule sets that are time-consuming to create and maintain. This thesis presents the design, implementation, and evaluation of a hybrid bank transaction reconciliation algorithm integrated into "Microsoft Dynamics Business Central“ system’s "Bankfeed" module. The core contribution is an algorithm‐selection meta‐model that, leveraging historical matching data, dynamically chooses the optimal reconciliation method – whether rule‐based or machine-learning-based, for each incoming transaction. During the analysis part, we identified the machine-learning model variant that most accurately predicts the best matching approach, measured its precision and coverage, and introduced weighting schemes to balance accuracy against computational cost. Finally, we report on a field trial assessing practical performance and statistically validate the improvements in matching accuracy and operational efficiency compared to the legacy rule-based system. This hybrid approach hopes to significantly reduce manual configuration effort and administration cost, while boosting the day-to-day effectiveness of accounting staff in ERP environments. |
