Multi-Source Models for Civil Unrest Forecasting.

Area of Science:

• Social Science
• Computational Social Science
• Data Science

Background:

• Civil unrest events, including protests and riots, can cause significant societal changes.
• Accurate detection and forecasting of these events are vital for policymakers and social scientists.
• Existing methods may not fully capture the complex dynamics leading to civil unrest.

Purpose of the Study:

• To develop and evaluate a multi-source data approach for forecasting daily civil unrest events.
• To identify the most informative data sources for predicting civil unrest in six Latin American countries.
• To provide insights into the drivers of civil unrest through data analysis.

Main Methods:

• Utilized logistic regression with Lasso for feature selection from diverse datasets.
• Integrated data from social media (Twitter, blogs), news, Tor network requests, political event databases, and exchange rates.
• Evaluated model performance using F1-scores against a Gold Standard Report (GSR).

Main Results:

• Forecasting models achieved F1-scores ranging from 0.68 to 0.95, demonstrating high efficacy.
• Social media and news data were found to be more predictive than political event databases.
• An ablation study confirmed the significant contribution of social media and news to prediction accuracy.

Conclusions:

• A multi-source data approach significantly enhances the prediction of civil unrest events.
• Social media and news are key indicators for forecasting civil unrest.
• Further research can refine models to manage the increased variability introduced by social media data.