Enhancing Stock Market Prediction With Hybrid Deep Learning: Integrating LSTM, Transformer Attention, Federated Learning, and Sentiment Analysis

Accurate stock market prediction remains a critical yet challenging task due to the highly non-linear, volatile, and sentiment-driven nature of financial markets. In this paper, we present a hybrid deep learning framework that integrates long-short-term memory (LSTM) networks with Transformer-based attention mechanisms, sentiment analysis from financial news, and a privacy-preserving Federated Learning (FL) strategy. First, we benchmark traditional forecasting approaches, including ARIMA, SARIMAX, Prophet, Random Forest, and Support Vector Regression, against the baseline LSTM models. Our results show that LSTMs consistently outperform conventional methods in capturing temporal dependencies. To further enhance predictive accuracy, we incorporate Transformer attention to improve long-range dependency modeling and apply sentiment analysis using FinBERT-tone to embed market sentiment signals into the model. Finally, we simulate a Federated Learning environment, enabling decentralized model training without sharing raw financial data, thus addressing privacy concerns in the financial domain. Experimental results in ten major technology companies (Tesla, Apple, Amazon, Microsoft, Google, etc.) demonstrate that our hybrid model achieves superior short-term forecasting performance, with an average R²variance score of 0.91 across ten major technology companies and a trend precision of 65.36%, demonstrating strong prediction performance for short-term stock forecasting. These findings highlight the potential of combining deep sequential models, attention mechanisms, and privacy-sensitive training strategies for robust and secure stock market forecasting.