An Ensemble Model for Financial Statement Fraud Detection

Abstract

Fraudulent financial statements are deliberate furnishing and/or reporting incorrect statistics, and this has become a major economic and social concern as the global market is witnessing an upsurge in financial accounting fraud, costing businesses billions of dollars a year. Identifying companies that manipulate financial statements remains a challenge for auditors, as fraud strategies have become increasingly sophisticated over the years. We evaluate machine learning techniques for financial statement fraud detection, particularly a powerful ensemble technique, the XGBoost algorithm, that help to identify fraud on a set of sample companies drawn from the MENA region. The issue of the class imbalance in the dataset is addressed by applying the SMOTE algorithm. We found that XGBoost algorithm outperformed other algorithms in this study: Logistic Regression (LR), Decision Tree (DT), Vector Machine Support (SVM), Adaboost, and RandomForest. The XGBoost algorithm is then optimised to obtain the optimum performance.

Predicting accounting fraud using imbalanced ensemble learning classifiers – evidence from China

“…Table 5 reports 3456 fraudulent firm‐year observations spotted from 1998 to 2017. Closely associated with the past literature (Bao et al, 2020; Khedr et al, 2021; Sun et al, 2021), an insignificant proportion of fraudulent observations exists in our collected dataset, generally at an estimated 10% of all companies annually. Therefore, our dataset is highly imbalanced, and the paucity of fraudulent instances underscores the continuing challenge for our research in forecasting accounting fraud.…”

“…An example is the work of Bao et al (2020) which applied two imbalanced classifiers (RUSBoost and SVM‐FK) in predicting financial statement fraud for publicly traded US companies. Khedr et al (2021) also utilised the SMOTE algorithm to avoid fraud data imbalance problems.…”