Fine-Tuned Autoencoder Neural Network for Anomaly Detection in Accounting Transactions

Abstract

Anomaly detection in accounting transactions plays a crucial role in identifying irregularities that may signal fraud, errors, or unusual financial behavior. Traditional rule-based and statistical methods often struggle to detect complex and hidden patterns in large-scale financial datasets. This paper presents a fine-tuned Autoencoder Neural Network for detecting anomalies in structured accounting records. The model processes feature such as date, account type, debit, credit, transaction category, and payment method. Preprocessing includes handling missing values, encoding categorical data, and extracting temporal features. The Autoencoder architecture was optimized using multiple hidden layers and dropout regularization to prevent overfitting. Reconstruction errors were used to determine anomaly scores, with a dynamic threshold set at the 98th percentile. Experimental results show that the model accurately distinguishes normal and anomalous transactions, identifying 2,000 outliers from a total of 100,000 records. Additional analysis indicates that anomalies often occur during weekends or holidays and involve unusual payment methods. These findings demonstrate the potential of the fine-tuned Autoencoder as a scalable and intelligent anomaly detection framework to support auditors and financial analysts in proactive fraud prevention.