Hannes Mueller (IAE-CSIC and BSE) and Christopher Rauh (University of Cambridge) have relaunched conflictforecast.org, a site that provides forecasts for outbreaks of violence and escalations into armed conflict.
Their method has benefitted from more than 8 years of research, published in the American Political Science Review, the Journal of the European Economics Association and International Interactions. The webpage now includes forecasts of conflict risk for more than 65,000 locations in the world. Users of the webpage can now switch between three different views:
- The conflict view shows outbreak risk for countries without internal armed conflict and escalation risks for countries with an ongoing conflict.
- The prevention view allows users to focus on our forecasts for countries without an ongoing armed conflict.
- Most importantly, the researchers have now added a sub-national view that shows outbreak and conflict risks for grid-cells all around the world.
As the world watches the horrors of the war in Ukraine, the motivation to prevent more wars should be as high as ever. Conflictforecast.org was developed with the motivation to help prevent the outbreak of new internal armed conflicts (civil wars). All too often attention only shifts to a country once violence has already broken out. This is too late. The researchers hope that by measuring escalating risks and publicizing them, preventive policies can be improved and tested. Tracking risks before the first outbreak of conflict is particularly important to the researchers. This prevention view, which disregards the most violent ongoing armed conflicts, allows users to enter into a mindset of preventive effort in which policy efforts don't need to rely on armed interventions.
A new technology, developed for the web page, allows the prediction of outbreaks and continuation of violence at the sub-national level, the so-called grid-cell level, at an unprecedented level of precision. The method divides the world's surface into more than 65,000 squares and tracks the conflict risk for these locations in monthly updates. By combining natural language processing, location detection and supervised machine learning the authors exploit the information contained in millions of news articles and connect it to the grid-cells. This allows the researchers to use both conflict dynamics and news information when forecasting conflict.
The conflict view mixes the outcomes of two models. For countries without an ongoing conflict it tracks the risk of an outbreak of an armed conflict. The method for this type of forecast has been published in the article The Hard Problem of Prediction for Conflict Prevention, forthcoming in the Journal of the European Economic Association. For countries with an armed conflict the project tracks the risk of a conflict escalation. This method has been developed for a conflict forecasting competition organized by the VIEWS project and has outperformed most other methods on the trained task (minimizing the MSE).
Forecasts are updated regularly and can be downloaded easily from the website: conflictforecast.org.
See also the team's paper, "Reading Between the Lines: Prediction of Political Violence Using Newspaper Text" in American Political Science Review and "Can machine learning help policymakers detect conflict?" (a non-technical summary in the BSE Focus).
