Extracting Alpha from the News Cycle
This documentation details a machine learning research project conducted in 2026 to evaluate how global news sentiment and economic calendars predict currency and commodity price movements. Researchers utilized the GDELT Global Knowledge Graph and economic event data to engineer fifteen unique features, testing them across 72 experimental configurations using gradient-boosted models like CatBoost and XGBoost. The study found that USDJPY was the most responsive instrument to news signals, with the highest-performing models achieving a Sharpe ratio of +2.653. Results indicated that a 24-hour prediction horizon most effectively captures the market drift following major news announcements. Ultimately, incorporating these alternative data sources provided a 32% performance boost over traditional technical baselines, proving that free news archives can offer a significant trading edge.