Does Global News Tone Add Predictive Value for Market Stress? Out-of-Sample Evidence from GDELT and the VIX

Introduction: This study examines whether globally aggregated news tone adds stable out-of-sample information for forecasting U.S. market stress beyond observable financial-market conditions.

Methods: Daily GDELT 2.0 Global Knowledge Graph indicators were aligned with the spot Cboe VIX, S&P 500 returns, and the 10-year U.S. Treasury yield from July 2016 to July 2026. Two leakage-safe outcomes were modeled: (1) a next-trading-day VIX jump exceeding the lagged 90th percentile of daily VIX changes over the preceding 252 trading days, and (2) entry within five trading days into a high-VIX state defined by the lagged 75th percentile of the VIX level. Logistic regression and XGBoost were compared using chronological holdout testing, expanding-window validation, probability recalibration, and 2,000-replication moving-block bootstrap inference.

Results: For next-day stress, adding news reduced logistic ROC-AUC from 0.634 to 0.595, but increased XGBoost ROC-AUC from 0.528 to 0.571 (Δ = +0.043, 95% CI: −0.000 to +0.089; p = 0.051). This gain was marginal and did not persist across historical validation windows. For five-day state entry, news reduced logistic ROC-AUC from 0.655 to 0.562 (Δ = −0.093, 95% CI: −0.190 to −0.008; p = 0.029) and reduced XGBoost ROC-AUC from 0.647 to 0.617 (Δ = −0.030; p = 0.375). Direct tone effects and tone-by-regime interactions were not robust.

Conclusions: Broad global news tone does not provide a stable stand-alone early-warning signal. Its only suggestive incremental value appears in nonlinear next-day VIX-jump prediction, while five-day regime-transition forecasts are better explained by prevailing volatility and recent market dynamics. Risk managers should therefore use global news tone only as a supplementary, dynamically validated input.