An artificial intelligence application of backpropagation neural networks to simulate accountants' assessments of internal control systems using COSO guidelines.

摘要

The objective of this study was to explore a form of artificial intelligence, neural network modelling, to examine variables that are crucial to technological implementation in accounting settings. The experiment utilized an accounting framework, assessing internal controls under COSO guidelines.; The results of this experiment suggest that a neural network model can be developed such that the decision processes of external auditors in assessing internal controls can be reasonably modelled. A significant difference exists between the classification precision of network models (a) using one hidden layer as opposed to two hidden layers, (b) between models with differing configurations of neurons within the hidden layer(s) and (c) a regression model for certain conditions.; A study of the incorrect decisions made by a neural network indicates that while reliance on a network would result in making some incorrect decisions, the threat of over-relying on internal controls is not extremely high. By eliminating noise in the research instrument, a network can be modelled that will be able to predict a higher number of correct assessments than was possible with the full experimental model.; A sensitivity analysis revealed that none of the five COSO inputs individually has an extreme effect on the neural network's ability to make internal control assessments. Self assessments by auditors who rate themselves as experts reveals that their models have a higher prediction rate at assessing internal controls than does the network developed from responses of lower self-assessed experts. Analysis reveals that auditors are extremely conservative in their response, especially in assessing controls over compliance with rules and regulations. Neural networks that were developed using effectiveness of internal controls as an outcome measure, had higher accuracy predictions than networks developed using quality of internal controls as an outcome measure.; This research demonstrates the usefulness of applying a neural network paradigm in assessing the effectiveness of internal control systems.