Published Paper Details:

Stock market prediction plays a significant role in financial analytics and strategic investment decisions. Traditional prediction models are often centralized, raising critical concerns regarding data privacy, ownership, and security--especially in regulated industries such as finance. This paper presents a novel approach that applies Federated Learning (FL), a decentralized machine learning technique, to stock price forecasting while preserving privacy. In our approach, we use historical stock data of Tesla Inc. (TSLA) and simulate multiple clients training Long Short-Term Memory (LSTM) models locally. The trained models are then aggregated by a central server over several rounds to form a global predictive model without exposing any raw data. A hybrid model strategy--combining federated updates with pre-trained centralized weights--is also implemented to improve training efficiency. Our system demonstrates an RMSE of approximately 34.72 and a MAPE of 6.8%, balancing accuracy with confidentiality. This study shows that Federated Learning is a feasible and effective solution for privacy-aware stock prediction in the financial sector.