Sentiment-Driven Market Microstructure Analysis of Digital Asset Trading Platforms

This study quantifies how exogenous sentiment innovations propagate through market microstructure across fragmented digital asset trading platforms using a multi-venue, high-frequency panel. The dataset spans 180 consecutive trading days, covering 5 venues, 12 liquid assets, 46.3 million trades, and 118.7 million top-of-book snapshots, aligned to a latency-aware event-time grid with Δ ∈ [1,15] minutes. Calibrated sentiment probabilities achieve macro-F1 between 0.792 and 0.851 across topical classes and expected calibration error between 0.029 and 0.044. Local-projection estimates show immediate market-quality deterioration after sentiment shocks: effective spreads widen by 0.214 at h=1, 0.162 at h=5, and 0.061 at h=15 (standardized units), while depth within ±1 bp contracts by -0.187 at h=1 and -0.141 at h=5. Realized volatility increases by 0.176 at h=1 and 0.132 at h=5, whereas order-flow imbalance rises by 0.093 at h=1 and decays to 0.028 by h=15. Negative sentiment produces stronger effects than positive sentiment, with a spread asymmetry of 0.083 and a volatility asymmetry of 0.071 at h=5. Regime conditioning indicates nonlinear amplification under thin-depth states, where the spread response at h=5 rises to 0.251 and depth contracts to -0.236. Price discovery reallocates toward low-latency venues as information shares shift by +4.8 pp and +3.2 pp to the two leading centralized exchanges under negative sentiment, while aggregators lose -4.2 pp and -5.2 pp. Out-of-sample evaluation using rolling 30-day training and 7-day testing windows shows sentiment-augmented models reduce RMSE for effective spread from 0.912 to 0.854 and for realized volatility from 0.936 to 0.895, with larger gains during stressed regimes. Placebo timestamp shifts of +30 minutes collapse the h=5 spread effect from 0.162 to 0.019, supporting temporal precedence and a shock-transmission mechanism mediated by liquidity provision and venue heterogeneity.