Does Investor Sentiment Predict Market Volatility? A Regime-Based Analysis Using Social Media Data

This paper examines the relationship between investor sentiment and stock market volatility using a regime-based empirical framework. The analysis draws on 367,665 tweets referencing 101 major global stocks over the period January 2017 to December 2018. Sentiment volatility is constructed as the mean intra-day standard deviation of tweet-level polarity scores aggregated first at the stock-day level, using two measures: an LSTM-based classifier and a TextBlob lexicon-based classifier. Market volatility is measured as a 5-day rolling standard deviation of mean absolute daily returns. Pearson correlation tests yield no significant linear relationship between sentiment volatility and market volatility for either measure (LSTM: r = −0.153, p = 0.097; TextBlob: r = −0.085, p = 0.355). Granger causality tests across lags of one to five days similarly find no significant predictive relationship. Regime classification based on the 90th percentile of market volatility produces 12 stress-regime and 107 normal-regime daily observations. Mann-Whitney tests confirm the regime partition is statistically valid (U = 0, p = 1.50 × 10⁻⁸). Within regimes, sentiment volatility is consistently lower during stress periods than normal periods for both measures (LSTM: d = −0.42; TextBlob: d = −0.46), suggesting that investor opinion on Twitter converges rather than diverges during elevated market volatility, though this difference does not reach conventional significance thresholds given the small stress-regime sample. Return differentials between high- and low-sentiment days are uniformly non-significant across all horizons and both measures. These findings indicate that daily-aggregated Twitter sentiment does not predict market volatility in this sample, but the consistent convergence pattern across both sentiment measures during stress regimes represents a directionally coherent finding that warrants investigation in larger samples.