Using Classification Algorithms for Smart Suggestions in Accounting Systems

Abstract

Accounting is a repetitive task and is mainly done manually. The repetitiveness makes it a suitable target for automation, however not much research has been done in the area yet. This thesis investigates how two di erent classification algorithms, Support Vector Machine with Stochastic Gradient Descent training and a Feed-Forward Neural Network, perform at classifying nancial transactions based on historical data in an accounting context. The classification algorithms show promising results but still does not outperform the existing implementation which is simple and deterministic. However, classi cation itself very much relies on the labels, i.e. how different users have accounted the transactions. As a response to this, we finally give a suggestion on how clustering might be used for the automation of accounting instead.