Self-Supervised Deep Learning for Cross-Market Anomaly Detection in Financial Systems

Abstract

 In the complex ecosystem of modern financial markets, anomaly detection plays a critical role in identifying irregular behaviors, fraudulent activities, and systemic risks. However, conventional supervised learning models rely heavily on labeled datasets, which are costly and time-consuming to obtain in real-world financial environments. This paper proposes a novel self-supervised deep learning framework for cross-market anomaly detection that effectively transfers learned representations across heterogeneous financial domains. The proposed framework integrates contrastive learning, graph-based relational modeling, and cross-domain feature alignment. Specifically, the method constructs proxy tasks using unlabeled data, such as temporal sequence reconstruction and context prediction, to pre-train a self-supervised encoder. This encoder is subsequently fine-tuned on downstream anomaly detection tasks across different financial markets, including equities, forex, and cryptocurrency datasets. Experimental results demonstrate significant improvements in anomaly detection performance compared to state-of-the-art baselines, especially under limited supervision. The proposed approach not only enhances generalization capability but also mitigates domain shift, leading to more robust detection of anomalous financial patterns across global markets.