Machine Learning–Based Detection of Financial Statement Frauds: Evidence from Firms Listed on the Indonesia Stock Exchange

Abstract

Financial statement fraud presents substantial threats to market integrity, investor trust, and economic stability, necessitating precise detection systems. Although machine learning techniques have been widely applied in fraud detection, empirical studies focusing on financial statement fraud in emerging markets using limited and highly imbalanced datasets remain relatively underexplored. This study proposes a machine learning framework to identify anomalies in companies' financial statements. It analyzes 914 firm-year observations derived from 119 firms listed on the Indonesia Stock Exchange (IDX), including 13 confirmed fraud firms, resulting in a highly imbalanced fraud-to-non-fraud ratio of approximately 1:8. Using 14 financial ratios as research variables, fraud detection models were developed based on Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), and Logistic Regression. The models were evaluated using accuracy, precision, recall, F1-score, Area Under the Receiver Operating Characteristic Curve (AUC-ROC), and Average Precision to better account for class imbalance conditions. The empirical results reveal that the RNN achieves a high recall rate and sensitivity in identifying fraudulent firms. The findings further indicate that while the RNN demonstrates strong recall performance and the LSTM captures temporal financial patterns effectively, Logistic Regression achieves the best overall classification performance in terms of accuracy, precision, and AUC-ROC. The findings provide practical implications for auditors, regulators, and investors by demonstrating that simpler statistical models may remain effective for financial statement fraud detection in emerging markets characterized by limited and imbalanced datasets.