Unsupervised-learning financial reconciliation: a robust, accurate approach inspired by machine translation

Abstract

Financial reconciliation (cross-checking independent sources of data) is a time-honored and widespread function in finance and audit. Its objectives are to ensure completeness, timeliness, and accuracy of recording of transactions. With the proliferation of data over recent decades, the ways in which reconciliation is approached have evolved. There are now a number of software products that promise to automate the process of detailed transaction matching. However, the 'state of the art' in these products is rules-based reconciliation. Rules-based systems in any domain tend to be brittle in that any change in the data streams causes the rules to have to be debugged and rebuilt, itself often a time-consuming process. And to make matters worse, this might be caused not just by data schema changes, but the contents of what is in the data fields themselves. Either of these can occur if a third party such as a bank changes its internal processes. In a sense, automated reconciliation is where machine translation was almost 70 years ago; IBM's 1954 Georgetown experiment approached Russian-to-English translation using rules, but it took another 4 decades for researchers in data-driven Artificial Intelligence (AI) to realize why machine translation did not initially live up to its promises and develop a truly robust methodology for machine translation based on unsupervised learning. It turns out that financial reconciliation can be cast as a machine translation problem based on unsupervised learning. To our knowledge, we are the first to propose this. Here, we demonstrate via experiments on real-life (albeit small-scale) financial data that this way of approaching the problem demonstrates promise in terms of accuracy, as well as solving the problem of lack of robustness inherent in rules-based approaches.