Financial Anomaly Detection Model Using Deep Neural Networks for Financial Statements in a Peruvian Organization

Resumen

Financial fraud remains a critical challenge globally, with estimated annual losses reaching $5.38 trillion. In Peru, the absence of advanced technological solutions has intensified this issue, leading to significant economic losses of up to $100,000 per incident. Methods like internal and external audits are considered traditional for fraud detection and have proven to be insufficient, identifying only 15% and 4% of fraud cases. To address these shortcomings, this study proposes a deep neural network (DNN) model to detect anomalies in financial statements, leveraging machine learning techniques to improve fraud detection capabilities and provide security in finances. The model analyzes structured financial data, detecting irregularities through feature engineering and anomaly detection techniques. A dataset of 356 financial records from a Peruvian company in the hydrocarbons sector for the years 2021 and 2022 is utilized. The model’s architecture consists of multiple densely connected layers optimized to capture nonlinear relationships within financial data. Furthermore, to compensate for the imbalance of classes, the Synthetic Minority Over-sampling Technique was used, enhancing the model’s ability to identify fraudulent patterns with greater accuracy. The proposed model demonstrates a substantial improvement over conventional machine learning techniques, achieving an accuracy of 80%, a recall of 80%, and an AUC of 86%, great performance. Additionally, the model efficiently processes financial data in a faster manner, making it suitable for real-time fraud detection applications in high-risk environments. This study underlines the prospect of deep learning to improve anomaly detection, strengthen financial transparency, and enhance risk management in Peruvian organizations and beyond.