Using machine learning to identify audit opinion shopping

We select a machine learning model to identify audit opinion shopping and analyze the factors driving the model. To this end, we use six models, namely random forest, gradient boosting decision tree, random undersampling boosting, logistic regression (LR), support vector machine and multilayer perceptron. Among them, LR outperforms the other models. Using game theory, we classify 58 features potentially affecting opinion shopping into audit object, audit subject and audit environment categories. LR is used to obtain each category's importance score. We find that audit object features play a crucial role in audit opinion shopping. We also validate and interpret important features. Finally, we use a model to predict audit collusion. Our paper extends the scope of machine learning to scientifically identify audit collusion risk and reveals important features of audit opinion shopping, which has implications for global audit practice.