머신러닝을 활용한 회계부정 탐지에 관한 탐색적 연구

초록

In this paper, we suggest a new accounting fraud prediction model using a machine learning approach. Using logistic regression to predict accounting fraud is subject to a class imbalance problem because of the small accounting fraud sample in the analysis. To overcome this problem, we utilize the RUSBoost approach based on ensemble learning. Unlike the logistic regression model, the RUSBoost approach is known as an approach to adjust the sample based on the distribution of minority groups and is less likely to subject to the class imbalance problem. By comparing the predictive ability of logistic regression and the RUSBoost approach on future accounting fraud, we find that the RUSBoost approach based on ensemble learning improves the accuracy of future accounting fraud prediction. Additionally, we find that improvement in the accuracy of the model is more pronounced for the accounting fraud with intention and the fraud with numbers in the financial statements. Lastly, we suggest some accounting variables that are closely related to future accounting fraud. In sum, we provide the first large-sample evidence of how the newly established machine learning approach improves the accuracy of fraud prediction, thereby being suggestive of the feasibility of machine learning approach research in the accounting area.

키워드