Investor Sentiment Index Based on Large Language Models and Its Predictive Analysis for the Shanghai Composite Index

Abstract

In the realm of financial markets, investor mood often intensifies or lessens the influence of new information on price changes. Recently, progress in Large Language Models like the Bidirectional Encoder Representations from Transformers (BERT) algorithm has shown better predictive abilities than traditional methods in foreseeing market trends, so this study builds an investor sentiment index using the BERT model to test its predictive effectiveness regarding the Shanghai Composite Index, which involves gathering post data from the East Money Stock Forum, after that, the BERT model was fine-tuned for sentiment classification and scoring, thus creating a daily sentiment time series, and the predictive power of the constructed index was assessed using two approaches, conventional econometric regression analysis and a multi-scale forecasting framework combining Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and Long Short-Term Memory(LSTM) network models. The empirical results proved that the BERT-based sentiment index could precisely grasp variations in investor sentiment, improving the explanatory and predictive ability for stock returns and the direction of market movement, demonstrating that Large Language Models are useful for sentiment analysis in financial text, offering a practical tool for investors and regulators to utilize sentiment-related insights.