Automated account reconciliation using probabilistic and statistical techniques

Purpose ‐ The purpose of the present paper is to investigate how methods from statistics, natural language processing, information theory, and other scientific fields can be brought to bear on account reconciliation. Practically, the goal is to reduce the number of labor hours it takes to complete a task which is widespread in various subfields of accounting including fraud investigation. Design/methodology/approach ‐ In this paper, the authors explore novel applications of data mining techniques from natural language processing and statistics to a particular account reconciliation problem. The authors are careful to avoid ad hoc heuristics and instead work with techniques that are theoretically justifiable; this means the techniques should be extensible (subject to appropriate modifications) to problem variants other than those that are explicitly considered here. The authors evaluate their techniques based on precision and recall ‐ standard measures from the field of information retrieval. Findings ‐ The paper finds that with careful tuning, it is possible to achieve near 100 percent precision (suggesting that the technique is highly accurate compared with an expert human reconciliation clerk) and close to 100 percent recall. Originality/value ‐ The current approach, unlike many previous approaches, looks to general principles of information theory rather than relying on heuristics which may work for one problem but not another. This approach is therefore highly general, and would apply to virtually any kind of accounting data (including even data where transaction descriptions are in a language other than English).