The Study on Auditor''s Internal Control Judgment Behavior - An Application of Neural Network

• Main Authors: Wu,Meng-Da, 吳孟達
• Other Authors: Chen,Kuo-Tay
• Format: Others
• Language: zh-TW
• Published: 1996
• Online Access: http://ndltd.ncl.edu.tw/handle/51247972721150876874

碩士 === 國立臺灣大學 === 會計學研究所 === 84 === To express proper audit opinion, auditors to a large extent must rely on their professional judgment while collecting and evaluating audit evidences. Yet, as judgment exercised in other professions, audit judgment itself lacks clear and objective criteria. Its process still remains largely unknown. Therefore, behavioral accounting studies often selected audit judgment as research topics and applied statistical techniques to analyze auditors'' judgment policy. This study, using neural net as the tool, attempts to capture the internal rules employed by auditors as they try to form judgment, and compares the judgment policy captured through neural net with the results of previous studies. This study survey 42 auditors in Taiwan''s big 6 CPA firms using a fractional factorial design of simulated cases questionnaires. Each case is distinct on variational combination of six important internal control factors in purchasing cycle. Auditors are asked to express their judgment about the strength of internal control of each case. We then use these data as our learning and testing cases for back- propagation neural net to capture auditor''s decision policy. We find that network model with an architecture of one hidden layer and three processing elements in hidden layer has the best prediction capability. Moreover, if the hidden layer is removed, the net model can not converge. This means that variables have interactive effects and representing auditors'' judgment policy by a linear model will result in significant prediction error. By analyzing the significance of these six cues, we find that auditors regard segregation of duties as the most important factor while payment control is second. With respect to prediction capability, the best model has a 88.68% accuracy rate when applied to the test cases.