Federated Learning and Privacy-Preserving Stock Prediction: Secure Collaborative Learning Across Financial Institutions

Abstract

The application of machine learning to stock prediction has demonstrated remarkable advances, with models such as the Stock State Space Graph achieving state-of-the-art performance through the integration of state space modeling and graph-based relational reasoning. However, the training of these models typically requires centralized access to large volumes of sensitive financial data, raising substantial privacy concerns and regulatory challenges. Financial institutions hold valuable proprietary trading data that could improve prediction models if shared, yet competitive sensitivities and data protection regulations such as GDPR and CCPA prohibit direct data sharing. This paper proposes a federated learning framework for privacy-preserving stock prediction that enables collaborative model training across multiple financial institutions without requiring raw data to leave individual institutions. Our approach, Graph-enhanced Federated Stock Prediction (GFedStock), extends the S3G architecture to a federated learning setting where institutions jointly train a shared model while keeping their data local. We incorporate differential privacy mechanisms and secure aggregation protocols to provide formal privacy guarantees, and analyze the privacy-utility tradeoff inherent in privacy-preserving approaches. Through extensive experiments on multi-institutional market datasets, we demonstrate that federated learning achieves 94-97% of the accuracy of centralized training while preserving data privacy, with the graph-enhanced architecture capturing cross-institutional market structure information. Our work contributes to the growing field of privacy-preserving machine learning in finance, providing a principled framework for secure collaborative learning in the financial services industry.