Research on the Correlation between News Sentiment and Stock Price Volatility

Abstract

This study employs natural language processing techniques—including BERT-based sentiment analysis, LDA topic modeling, and LSTM time-series forecasting—to quantitatively analyze the relationship between news sentiment and stock price volatility in China's A-share market from 2020 to 2023. The dataset comprises 523,847 financial news articles from Sina Finance, Eastmoney, and other mainstream Chinese financial media, covering CSI 300 constituent stocks. A FinBERT model fine-tuned on 10,000 manually labeled Chinese financial news samples (accuracy: 87.3%) was used for sentiment classification. The results reveal a significant positive correlation between sentiment scores and stock price volatility (Pearson's r = 0.048, p < 0.01), while news volume exhibits a weak negative correlation with volatility (Spearman's ρ = -0.028, p < 0.05). Negative news exerts a stronger market impact than positive news, demonstrating clear asymmetric effects. Deep learning models significantly outperform traditional approaches: LSTM achieves a classification accuracy of 73.5% (F1 = 0.72), compared to 52.3% (F1 = 0.49) for linear regression. The study provides theoretical insights into behavioral finance and information asymmetry theory, while offering practical data-driven support for investor decision-making, corporate reputation management, and regulatory risk warning systems.