Application of an LSTM-Transformer Hybrid Model Incorporating Public Sentiment in Predicting CSI 300 Stock Prices

Abstract

The short-term price changes in the stock market are influenced by many factors, so it is difficult to predict. In this paper, the daily trading data of CSI 300 index from 2016 to 2025 and the emotional text of stock forum are used as inputs to construct the LSTM-Transformer hybrid model. It uses CFSD China Financial Emotion Dictionary to quantify emotions, and evaluates the predictive advantages of emotional information through ablation experiments. The results show that the average MAE of the mixed model is 77.59, RMSE is 107.81, and r is 0.9218, which is better than the LSTM and Transformer models alone. After adding emotional characteristics, the average MAE decreased by 21.6%, RMSE decreased by 15.4%, and R decreased from 0.8715 to 0.9146, and remained stable in the 1-5-day forecast. This study provides a practical solution to predict, mix sensory data with mixed models.