A Predicting Model For Accounting Fraud Based On Ensemble Learning

Accounting fraud, usually difficult to detect, can cause significant harm to stakeholders and serious damage to the market. Effective methods of accounting fraud detection are needed for the prevention and governance of accounting fraud. In this study, we develop a novel accounting fraud prediction model using XGBoost, a powerful ensemble learning approach. We respectively select 12 financial ratios, 28 raw accounting numbers and 99 raw accounting numbers available from Chinese listed firms' financial statements, as the model input. To assess the performance of fraud prediction models, we select two evaluation metrics - AUC and NDCG@k, and two benchmark models - the Dechow et al. (2011) logistic regression model based on financial ratios, and the Bao et al. (2020) AdaBoost model based on raw accounting numbers. Results show that: 1) our XGBoost-based prediction model outperforms two benchmark models by a large margin whatever model inputs and evaluation metrics; 2) the XGBoost-based prediction model with raw accounting numbers input outperforms the one with financial ratios input; 3) the XGoost-based prediction model with 99 raw accounting numbers input outperforms the one with 28 raw accounting numbers input.