A Comparative Study of XGBoost and Artificial Neural Networks for Earnings Management Prediction

The present study was conducted to compare the accuracies for ANN and XGBoost algorithms on predicting earnings management in listed companies of Tehran Stock Exchange. Earnings management is one way for managers to mislead their stakeholders, which can result in financial losses; therefore, accurate detection methods are important for earnings management and beyond statistical models. The present study used ۲۰۱۶–۲۰۲۱ quarterly financial data from ۱۰۳ publicly traded companies in basic metals, automotive, chemical, food and pharmaceutical producing industries (۵۰۷۶ year−firm). Discretionary accruals were used and calculated by the Kasznik model to capture earnings management and split them into three groups: Increasing Accruals (+۱), Decreasing Accruals (-۱), and Near-Zero Accrual (۰). A Confusion Matrix was used to perform the model evaluation. The results showed that the XGBoost algorithm is significantly superior to the ANN with an overall accuracy of ۹۸.۴%. With very few errors, all earnings management categories XGBoost had near to perfect results. In contrast, ANN demonstrated significant weaknesses leading to an overall accuracy of ۶۳.۱%. The study found that the model performance of XGBoost can further predict earning management with more accuracy thereby securing a process for financial institutions. Inspired by the relatively rare occurrence of applying XGBoost model, used for trilateral classification of increasing, decreasing and near-zero earnings management in emerging markets including Tehran Stock Exchange. This research contributes significantly to the literature by demonstrating the superior predictive power of ensemble learning methods over traditional neural networks in detecting financial misconduct, which offers a robust, high-accuracy tool for regulators and investors to enhance market transparency and reduce financial risk.

کلیدواژه ها:

Artificial Neural Network (ANN) ، Earnings management ، Kasznik Model ، Machine learning ، XGBoost