Federated Deep Reinforcement Learning For Privacy-Preserving Sentiment-Driven Stock Market Forecasting

Abstract

In this paper we propose a new FL + DRL combination model, which can reinforce a sentiment-based prediction model in the stock market that offers improved privacy and forecasting. Sentiment signals are trained in the proposed method by extracting the signals in financial news, social media, and earning calls transcripts by using advanced natural language processing (NLP) models, and the trained signals will form the basis of how the policy is to be learned in DRL. The use of FL enables financial institutions in different countries to directly share the collective financial forecasting model without access to the sensitivity of local data to comply with regulations and data privacy. The reinforcement learning aspect of the system allows learning the dynamic and non-stationary dynamics of the financial markets, i.e. is to keep improving the trading strategies over time. Compared to the centralized and standalone models, experimental analysis conducted across various market indices and asset classes proves that the offered federated DRL architecture is much more effective in terms of predictive performance and its sparsity to adversarial noise in sentiment input. Also, ablation experiments prove the beneficial effects of sentiment integration and federated updates on portfolio performance over time. This study is timely as it lies on the convergence of AI, finance, and privacy and offers a scalable technology resulting in joint financial intelligence without threatening confidential information, and establishes a future of safe, real-time decision making in high frequency trading markets.