A semi-heterogeneous ensemble forecasting method for stock returns based on sentiment analysis

Abstract

With the growing influence of investor sentiment on market dynamics, sentiment analysis has emerged as an effective tool for enhancing financial forecasting models. This study proposes a diversity-enhanced semi-heterogeneous ensemble forecasting framework that integrates sentiment analysis into the forecasting of stock index returns. A supervised stock market sentiment index set is constructed, in which prior knowledge regarding term importance is integrated into the data augmentation process. This enables higher weights to be assigned to sentiment-related terms with superior predictive capacity, thereby allowing the model to prioritize more informative features and enhance its forecasting performance. A series of diverse base models are generated through the integration of multiple attention-PCA techniques and forecasting algorithms based on variable perturbation strategies. These base models are subsequently combined through a suite of ensemble strategies, forming a semi-heterogeneous ensemble model for forecasting S&P 500 returns. The experiment results demonstrate that the proposed approaches significantly outperform benchmark methods, with notable improvements in both accuracy and diversity.