Forecasting realized volatility of state-level stock markets of the United States: The role of sentiment

This paper examines whether state-level sentiment innovations improve forecasts of realized volatility in U.S. regional stock markets. Using 5-minute intraday data for 50 U.S. states and a daily Twitter-based state-level sentiment index over the period August 2011 to August 2024, we evaluate the incremental predictive content of changes in local sentiment within a Heterogeneous Autoregressive realized-volatility framework. To isolate the role of sentiment, the forecasting models also control for intraday distributional information, including realized skewness and realized kurtosis. Out-of-sample forecasts are generated using a rolling-window design at daily, weekly, and monthly horizons, with predictive accuracy assessed through formal forecast comparison tests. The results show that sentiment innovations contain useful, though heterogeneous, forecasting information. While their average in-sample effect is limited, changes in sentiment improve out-of-sample volatility forecasts for a non-negligible subset of states, especially at short horizons and when combined with higher-moment controls. Additional evidence indicates that local sentiment retains predictive content beyond aggregate sentiment and that its forecasting relevance is linked to state-level economic conditions. Overall, the findings suggest that sentiment is a distinct, geographically heterogeneous, and episodic source of information for forecasting state-level stock market volatility.