Impact of investor sentiment on stock market using sentiment estimation and multi-head attention LSTM network

In this study, we explore the impact of investor sentiment on stock market dynamics through stock closing price predictions. We propose a Multi-Head Attention Long Short-Term Memory Network (MA-LSTM) designed to predict stock closing prices by integrating both stock features and investor sentiment. By merging the capabilities of Long Short-Term Memory (LSTM) and Multi-Head Attention mechanisms, MA-LSTM adeptly captures the temporal dependencies concealed within stock market data and investor sentiment features. Investor sentiment is estimated using a Bidirectional Encoder Representations from Transformers (BERT) model, based on investor messages collected from social media platforms. To enhance sentiment estimation accuracy, we conduct further pre-training of the BERT model in the stock market domain. We combine investor sentiment with stock price data and feed it into the MA-LSTM model for predicting the closing prices of prominent stocks such as Apple and the SPDR S&P 500 ETF. The experimental results demonstrate the superiority of the proposed method over the traditional LSTM model, regardless of the inclusion of sentiment features. Particularly the MA-LSTM model with sentiment features has good effectiveness. It’s evident that incorporating sentiment features enhances the forecasting performance of stock closing prices.