Historical Data Based Sentiment-Driven Stock Market Prediction Using a Hybrid BERT-LSTM Model

Forecasting patterns and trends in the stock market is a crucial activity that needs to be accomplished in order to enhance the investment plans of involved investors, but it is not an easy task to perform given the constantly fluctuating and highly dynamic nature of the markets. One major drawback of traditional forecasting models is the inability to incorporate sentiment analysis into the historical price data. This study addresses this issue by proposing a novel hybrid model that combines the strengths of Bidirectional Encoder Representative from Transformer and Long Short Term Memory networks. Bidirectional Encoder Representative from Transformer analyzes and converts sentiment shared in financial news and on social media and to capture subtle emotional data from textual information. While, Long Short Term Memory network detects sequential pattern and the trends of the historical stock prices which provides a better understanding of the temporal data. The combination of these two components in the proposed framework enhances forecast accuracy and provides in-depth insights into market movements. As demonstrated by the experimental data, the Bidirectional Encoder Representative from Transformer and Long Short Term Memory model outperforms conventional approaches in stock market trend forecasting, achieving an MAE of 15, MSE of 0.9, RMSE of 3, R2 of 0.9, and Mean Absolute Percentage Error of 1.2.