Advanced Stock Index Prediction Using Optimized LSTM Model Incorporating Sentiment Analysis

Abstract

Stock index prediction is essential for financial market analysis and investment decisions, but traditional methods like ARIMA and GARCH often fail to deliver accurate results due to the complex and volatile nature of the stock market. Advances in deep learning, such as LSTM and GRU models, have shown promise but still face limitations in fully capturing market sentiment. The aim of this study is to develop a novel stock index prediction model that integrates GRU and LSTM to enhance predictive accuracy. The study also incorporates BERT-based sentiment analysis to better capture investor emotions and improve prediction performance. The proposed model utilizes data from the Stockholder Sentiment Dataset and Stock Index Trading Dataset, sourced from the Shanghai and Shenzhen Stock Exchanges. BERT is used to calculate sentiment scores from public news data, and an optimized combination of GRU and LSTM is employed for time series prediction. Comparative experiments were conducted against baseline LSTM and GRU models using metrics such as accuracy and loss.

The proposed model achieved an accuracy of 88.92% and a loss of 0.18 after 30 epochs, outperforming the basic LSTM (83.21% accuracy, 0.22 loss) and GRU models (85.34% accuracy, 0.20 loss). This demonstrates the superior performance of the model in stock index prediction tasks. This study introduces a more effective stock index prediction model by combining advanced deep learning methods and sentiment analysis. The proposed model offers valuable insights for improving investment strategies and risk management in financial markets, providing a foundation for further research in this field.