Federated Learning for Financial Forecasting: A Privacy-Preserving Approach to Decentralized Data Sharin

Abstract

This project examines the use of federated learning for financial forecasting, which focuses on a better prediction with privacy. Data aggregation in centralized models can break confidentiality, notably in finance. Our study offers a federated learning (FL) paradigm utilizing long short-term memory (LSTM) networks whereby diverse financial institutions collectively train strong forecasting models without data sharing. We employed NASDAQ-100 and S&P 500 datasets and utilized a differentially private LSTM network leveraging secure multiparty computing. The data reveal that performing an averaging federated (FedAvg) model was much superior to centralized and decentralized models with lower MAE and RMSE. The model's R2values of 0.92 show its ability to capture the market's complexity and perform well. This framework secures privacy and enables scalability for realtime financial forecasting. According to our findings, federated learning has the ability to substantially impact the banking industry and give an accurate and secure alternative to the existing approaches. Future studies will aim at including sophisticated privacy-preserving approaches and increasing model applications across varied financial datasets.