Belief Networks for Expert System Development in Auditing

This study examines the use of a belief network based expert system for an auditing task—financial distress evaluation for banks. A belief network uses probability measures to store important dependencies across variables of interest in a problem domain, and makes inferences based on observed evidence using probability calculus. This paper discusses how belief network structures can be constructed, and used to assist auditor's in making appropriate recommendations regarding the financial health of a bank under audit. The ability of a belief network to make reliable predictions depends on how well the network structure reflects the underlying dependencies across variables in the problem domain (e.g. financial ratios and the financial health of a bank). The first part of this study illustrates how a computer program developed by the authors can be used to generate and evaluate different feasible belief network structures based on historical data. The program uses an information‐theoretic measure to compare the alternative structures. The ability of the program to identify existing dependencies across variables is demonstrated by using it to reconstruct a known network structure from simulated data. Next, the program is used on a database of twelve important bank financial ratios over a three‐year period. The predictive ratios identified by the program reflect important areas of a bank's health, such as loan quality, efficiency, profitability and capital adequacy. Finally, a belief revision mechanism is encoded for the belief network structure identified earlier, and is used to illustrate how it can assist auditors in making recommendations about financial health based on a bank's critical financial ratios. The probability estimates provided by the system are validated using data on banks not used in the network design stage, and are found to be reliable.