Beyond traditional GARCH models: integrating NLP-derived sentiment scores for enhanced volatility modeling

Abstract

Financial markets react dynamically to new information, influencing asset returns volatility. While traditional GARCH models effectively capture volatility clustering, they often overlook the role of exogenous factors, particularly sentiment. This study extends the conventional GARCH framework by incorporating a sentiment driven exogenous component to enhance volatility forecasting. We develop a methodology that integrates sentiment extracted from financial news using FinBERT with contextualisation via GPT-4o mini, ensuring a more precise representation of market sentiment. Through simulations of the GARCH-X process, we demonstrate that the sentiment-enhanced GARCH-X model provides more accurate volatility estimates than the conventional GARCH model, as reflected in its lower mean squared errors and AIC values. Fitting our GARCH-X model to S\&P500 and Tesla log returns, we find that exogenous sentiment scores are statistically significant. By integrating sentiment analysis into the volatility estimation, the GARCH-X model captures market dynamics more comprehensively, enhancing the predictive power for financial time series. This improvement in volatility estimation could lead to better risk management strategies and optimized portfolio allocations, ultimately contributing to more robust investment outcomes.