Return predictability via sentiment: individual, industry, or market?

This study investigates the hierarchical predictability of investor sentiment across individual, industry, and aggregate market levels, utilizing daily data for 6,482 U.S. stocks (1999–2019). We benchmark traditional econometric models against nonlinear machine learning techniques, including XGBoost and LSTM. Results indicate that while individual sentiment robustly forecasts specific stock returns (XGBoost R2≈ 0.37), aggregated industry-level sentiment—particularly in the Finance and Healthcare sectors—provides the strongest signals for broader market outcomes. Nonlinear models consistently outperform linear baselines, capturing complex spillover dynamics driven largely by mid-cap firms. These findings validate behavioral theories of sentiment-induced mispricing and offer actionable frameworks for algorithmic trading and systemic risk monitoring.