Trends in sentiment matter

 

The paper "Interpretable machine learning for macro alpha: A news sentiment case study" uses the Global Database of Events, Language and Tone (GDELT) with FinBERT, a bidirectional encoder representation from Transformers (BERT), to construct daily sentiment indices that incorporate time dispersion and event impacts. These indices are compared using an XGBoost classifier against logistic regression to predict major currency and bond futures. The study finds that sentiment trends are also useful for predicting asset returns, as shown by Shapley Additive Explanations (SHAP). Domain-specific natural language processing with interpretable ML can add value to any market model and generate macro alpha.