Real-Time Stock Market Sentiment Analysis Using Finbert and Lstm-Based Price Prediction Models

Abstract

With the growing importance of financial perception of market dynamics and forecasts for stock market prices has become an important field of intersection between the fields of natural language processing and deep learning. Current methods have not been successful at incorporating real-time sentiment into historical pricing information, leading to inaccurate forecasts in periods of market volatility. A comprehensive model by integrating the FinBERT-based contextual sentiment encoding and the stacked Bidirectional LSTM (BiLSTM) model with temporal self-attention and adaptive confidence threshold gating is proposed. The multivariate construction mechanism is designed to merge the financial sentiment and price data for the 66,480 pairs of aligned sentiment and price data in a sliding window fashion, where each pair involves a constituent of the S&P 500, such as AAPL, MSFT, TSLA, GOOGL, and AMZN. The proposed model shows that an average accuracy of 97.87%, MAE of 2.330, RMSE of 3.085, and $\mathbf{R}^{\mathbf{2}}$ of 0.9671 is obtained with all the stocks evaluated. The results show that the confidence-gated sentiment fusion model based on attention-weighted temporal pooling significantly outperforms the existing baseline models based on FinBERT-LSTM in real-time stock price prediction.