Robust Anomaly Detection in Financial Markets Using LSTM Autoencoders and Generative Adversarial Networks
Abstract
Anomalies in financial markets-characterized by sudden shifts in returns or volumes-can indicate systemic risk, structural breakpoints, or market manipulation. Detecting such events is critical for ensuring the resilience of trading systems, early- warning tools, and financial surveillance mechanisms. However, the absence of labeled anomaly data and reliance on high- frequency datasets often limit the practical deployment of sophisticated detection models.
In this study, we present a novel hybrid anomaly detection framework that operates effectively on widely available daily return and volume data. Our approach integrates a Long Short-Term Memory (LSTM) Autoencoder with a Generative Adversarial Network (GAN), capturing both temporal dependencies and distributional shifts in financial time series. To enhance precision in latent anomaly identification, we incorporate a One-Class SVM atop the LSTM-encoded representations. Additionally, we propose an artificial anomaly injection mechanism that simulates realistic market irregularities-such as price shocks and volume spikes-enabling quantitative evaluation in the absence of ground truth labels.
We conduct extensive experiments across six representative stock categories (e.g., indices, mega-cap, small-cap, high/low volatility, and penny stocks) and multiple macroeconomic regimes-including the Global Financial Crisis and the COVID-19 recovery. Our hybrid model consistently outperforms classical baselines (e.g., GARCH, Z-Score, One-Class SVM) in recall and F4-score, demonstrating robustness under both stable and turbulent conditions. Key contributions include: (1) a scalable, interpretable LSTMGAN hybrid framework tailored for anomaly detection on lowfrequency financial data, (2) a novel anomaly injection protocol for model validation, and (3) a systematic evaluation pipeline across diverse asset types and historical market regimes.
This study presents a practical and generalizable solution for anomaly detection in financial time series, rigorously evaluated to ensure reliability. It aims to bridge the gap between academic modelling and real-world deployment, particularly in data constrained environments.