Abstract
Traditional measures of national peacefulness, such as the Global Peace Index (GPI), suffer from significant reporting latency, often failing to capture rapid socio-political shifts. This study introduces a machine learning framework that utilizes the GDELT news database to "nowcast" and forecast GPI scores for 163 countries at a monthly resolution. By employing XGBoost and SHAP explainability, the research demonstrates that global news media serves as a high-frequency proxy for peacefulness, offering a robust early-warning tool for policymakers.
Introduction
The Global Peace Index (GPI) is a comprehensive measure of national stability, yet its annual publication cycle limits its utility for real-time crisis management. Abrupt events like civil unrest or terrorist attacks are often missed by the slow-moving institutional surveys that inform the index. To address this information gap, this research leverages Big Data to create a monthly estimation of peacefulness, transforming the voluminous stream of global news into a structured signal of national stability.
Methodology
The predictive pipeline involves several critical stages:
- Data Harmonization: Linear interpolation was used to convert annual GPI data into a synthetic monthly ground truth.
- Feature Engineering: Event counts were extracted from the GDELT database and categorized into 200 types based on the CAMEO (Conflict and Mediation Event Ontology) format.
- Predictive Modeling: Six models were evaluated, including Linear Regression, Elastic Net, and Support Vector Regression, with XGBoost emerging as the most effective.
- Rolling Window Training: A 72-month rolling window was utilized to account for political non-stationarity, predicting outcomes 1 to 6 months ahead.
- Explainability: SHAP (SHapley Additive exPlanation) values were applied to identify specific news events driving the peace scores, moving beyond "black box" predictions.
Key Insights
The study reveals that news media attention is a structured reflection of institutional peace. For instance, Iceland's peacefulness is characterized by "praise" and "endorsements," whereas the United States' score is influenced by "aerial weapons" and "judicial actions." However, the model can fall into a "Reactionary News" trap, where diplomatic efforts following a crisis (e.g., appeals for aid) are initially misinterpreted as positive stability markers. Furthermore, model performance naturally decays as the forecast horizon extends, and context-specific adjustments—such as shortening training windows for conflict zones—can improve accuracy.
Results
XGBoost consistently outperformed other models, achieving a Pearson Correlation greater than 0.8 for many countries, including the UK, Yemen, and Uganda. The Mean Absolute Percentage Error (MAPE) remained below 5% for high-performing models. Crucially, the monthly model successfully captured deteriorations in peace, such as the 2016 Belgium terrorist attacks and the 2014 Yemen Civil War, well before they appeared in official annual reports.
Limitations and Future Directions
While promising, the framework has limitations. Linear upsampling of the GPI is a simplification, and GDELT's reliance on English-language media introduces potential bias. Additionally, while SHAP identifies correlations, it does not establish causality. Future research should explore more complex stochastic models for ground truth estimation and integrate theoretical social interaction models to deepen the understanding of conflict drivers.
Summary updated