Machine learning in automating bank statement postings
Ojala, Juho (2018)
Ojala, Juho
Metropolia Ammattikorkeakoulu
2018
All rights reserved
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2018111817401
https://urn.fi/URN:NBN:fi:amk-2018111817401
Tiivistelmä
The thesis studies the possibility to use machine learning in bank statement ledger posting. It’s a feature that could potentially save a lot of manual work from accountants. The scope was restricted to determining ledger account instead of the full ledger posting. The goal was to build a proof of concept, rather than a production ready software. A secondary objective was to study potential tech choices for future use in the company for machine learning projects.
Creating a functional set of ledger posting rules for thousands of different companies with the ability to change over time is something that is very difficult and arduous to do using traditional programming techniques and provided a great opportunity to test machine learning concepts.
A proof of concept was built using a combination of machine learning techniques and some traditional programming fine-tuning it. This thesis covers the applied theory, materials and methods that were used, the detailed description of the solution that was built, an evaluation of the software performance and conclusions that can be drawn from those.
The results of the thesis indicate that some companies are more applicable for this kind of automatic processing than others and for production ready usage it is important to do better categorization of the companies. Overall the problem is more complex than initially expected and requires further development for production ready solution. Additionally, tech choices used were good and scaling computing power can be considered a requirement for effective machine learning development with larger amounts of data.
Creating a functional set of ledger posting rules for thousands of different companies with the ability to change over time is something that is very difficult and arduous to do using traditional programming techniques and provided a great opportunity to test machine learning concepts.
A proof of concept was built using a combination of machine learning techniques and some traditional programming fine-tuning it. This thesis covers the applied theory, materials and methods that were used, the detailed description of the solution that was built, an evaluation of the software performance and conclusions that can be drawn from those.
The results of the thesis indicate that some companies are more applicable for this kind of automatic processing than others and for production ready usage it is important to do better categorization of the companies. Overall the problem is more complex than initially expected and requires further development for production ready solution. Additionally, tech choices used were good and scaling computing power can be considered a requirement for effective machine learning development with larger amounts of data.