Wasserstein Generative Adversarial Network with Gradient Penalty and Threshold-Enhanced for Imbalanced Panel Data for Financial Fraud Detection

Abstract

Financial statement fraud detection is critical to maintaining trust among investors, regulators, and analysts. However, traditional audit procedures often fail to detect anomalies effectively because they occur infrequently but can result in significant economic losses. This study proposes an oversampling approach using a modified threshold in the Wasserstein Generative Adversarial Network with Gradient Penalty (WGANGP) to enhance synthetic data variance in financial fraud detection. The financial data were collected from the financial reports of companies listed on the Indonesia Stock Exchange and were labelled according to the Balanced Scorecard framework into four categories: normal, alarm, risky, and fraud. Given the severe class imbalance, this study introduces a WGANGP model with threshold optimisation in the generator and a gradient penalty to generate high-quality synthetic samples. This study conducted general and per-entity oversampling scenarios and evaluated them using Euclidean distance, Wasserstein distance, and classification metrics. In Scenario 1, the Generative Adversarial Network (GAN) outperformed the Synthetic Minority Oversampling Technique (SMOTE) and vice versa in Scenario 2. However, in financial fraud detection, the WGANGP with enhanced thresholding improved the F1-score by 13% to 17% compared to SMOTE and five GAN-based models across thirteen classification models, including traditional, machine learning, and deep learning models. This finding suggests that optimising the threshold in WGANGP reduces variance and improves model performance. Furthermore, generating synthetic data that is very similar to actual data may not necessarily improve classification; therefore, it is necessary to test how oversampling affects subsequent stages.