REAL-TIME ANOMALY DETECTION IN FINANCIAL TRANSACTIONS USING STREAMING ANALYTICS

Abstract

This study presents a hybrid deep learning framework for real-time anomaly detection in financial transactions through the combination of Long Short-Term Memory (LSTM) networks with Autoencoders (AE) deep learning models. The proposed LSTM model in this study is used to capture sequential dependencies and temporal patterns in transaction sequences, while the AE is used to identify deviations from normal behaviour through reconstruction errors. Then, the models were trained and evaluated using the Metaverse Financial Transactions Dataset which was acquired from Kaggle platform, and it is made up of a rich combination of transactional, behavioural and risk-related attributes, including timestamps, sender and receiver addresses, transaction amounts and types, simulated IP prefixes, location regions, login frequency, session duration, pre-computed risk scores, and labelled anomalies such as phishing and scam indicators. Preprocessing steps included data cleaning, standardization, one-hot encoding of categorical features, sequence construction for LSTM modelling, and feature engineering to capture behavioural trends. The hybrid LSTM-AE model achieved an accuracy of 96.4%, precision of 94.8%, recall of 95.7%, F1-score of 95.2%, and AUC-ROC of 0.983, outperforming the individual LSTM and AE models. These results demonstrate that integrating temporal modelling with reconstruction-based anomaly detection provides a robust framework for detecting fraudulent transactions in decentralized and high-velocity digital financial environments.