This study aims to examine whether sentiment polarity or public attention conveys informational value for Bitcoin return and volatility dynamics by separating human-generated signals from automation-driven amplification.
The analysis uses more than sixteen million Bitcoin-related tweets, classifies accounts into human-like and automation-like groups and constructs separate sentiment and attention indices. These indicators enter a multi-stage empirical framework comprising return-prediction models, GARCH-X volatility estimation and VAR-based return-attention dynamics across volatility regimes.
Polarity-based sentiment, hype, anxiety and divergence exhibit no predictive power for returns across all specifications. Public attention, however, significantly amplifies conditional variance and improves GARCH-X model performance while offering no directional content for returns. VAR and Granger causality show that attention reacts to price shocks rather than forecasting them. Automation-like accounts dominate the dataset and dilute polarity signals, whereas attention remains robust as a behavioural intensity measure.
The study demonstrates that attention, not textual polarity, drives short-horizon volatility in cryptocurrency markets and provides a refined empirical framework for modelling digitally mediated market behaviour.
