Audit opinion decision using artificial intelligence techniques: Empirical study of UK and Ireland

Abstract

The reliability and quality of a final decision regarding auditing opinion is a significant issue for auditors. The new field of data mining in auditing remains in its infancy and is increasingly explored through creating reliable and effective auditing opinion classification models. Previous studies have called for more exploration that is needed of the individual classifier models and committee combiner methods in the auditing field. Particularly, previous studies have not yet investigated or applied data mining dynamic modelling and even they have not encouraged future studies to do search in dynamic modelling in auditing and accounting area. Thus, this thesis study investigates the ability of a classification tool to classify correct audit opinion and explores dynamic modelling. To the best of this researcher‟s knowledge, this is the first research that involves dynamic modelling research in auditing opinion. Two evaluation measurement parameters that have not been used in any previous auditing studies or any related area are used to evaluate performance accurately: Brier score and area under reliability diagram (AURD). This thesis aims to develop the ability performance of nine classifiers (support vector machines, artificial neural networks, K-nearest neighbour, decision trees, naïve Bayes network, logistic regression, linear discriminant analysis, boosting ensemble and a novel deep learning model), offering a classification tool for correct audit opinion. The empirical evaluation results indicate that for the four tested datasets, the deep learning model revealed superior ability in classifying the audit opinion accurately, outperforming all other models by obtaining highest values at all nine evaluation parameters. Subsequently, significance statistical testing revealed that the deep learning model has best ability to classify audit opinion correctly. Thereafter, the audit opinion model was enhanced by combining all nine individual classifiers to improve the accuracy of the audit opinion modelling according to six traditional committee modelling rules (Average, Weighted average, Median, Min; Max and Majority voting). Moreover, the Consensus combiner and Fuzzy logic combiner models were added to the committee modelling. The performance of each committee modelling technique was assessed individually, and subsequently their abilities to classify audit opinion correctly were compared to determine whether committee modelling can improve upon the accuracy performance of individual classifiers. Consensus model showed superior ability to classify audit opinion correctly, and enhanced accuracy in audit opinion modelling compared with individual classifiers, which delivered the best evaluation measurement results over the four datasets, and the best statistical test results. The final contribution was developing of traditional dynamic modelling methods (nonlinear autoregressive exogenous and nonlinear autoregressive) and novel dynamic model (deep learning-LSTM), utilised to predict audit opinion in advance. These models were tested, and individual performance results being compared with the benchmark model result, which was a deep learning classifier that tested actual audit opinion data for the advanced year. Lastly, all dynamic modelling performances were compared using the benchmark classifier. Deep learning-LSTM had better performance in predicting audit opinion in advance compared with the other models, in terms of the best evaluation results.