Modeling and Analysis of Stock Price Prediction Incorporating Sentiment Attributes of Company Research Reports

Stock price prediction now sees a growing relies on advanced models to navigate the complexities of financial markets. While Transformer-based architectures excel at modeling temporal dependencies in stock data, they often overlook sentiment signals from professional sources such as analyst reports. The importance of high-quality input factors for enhancing prediction accuracy is often as significant as the algorithmic model itself. Addressing this gap, this paper introduces two novel sentiment features: ybscore and dybscore, designed to capture absolute sentiment levels and daily sentiment dynamics from analyst report ratings and rating changes, respectively. These features are integrated into two representative Transformer-based stock prediction models, Quantformer and Stockformer, to improve stock price forecasting and trading strategy performance. Empirical validation on 88 constituent stocks of the CSI 300 Index in the Chinese A-share market (2010-2024) demonstrates that the incorporation of these expert-derived sentiment features significantly enhances the models’ predictive capabilities and overall performance. This research underscores the potential value of systematically integrating structured sentiment data from professional analyses into stock trading strategies, offering insights for the development of more robust and informed investment approaches.