The Hard Problem of Prediction for Conflict Prevention
In this article, we propose a framework to tackle conflict prevention, an issue that has received interest in several policy areas. A key challenge of conflict forecasting for prevention is that outbreaks of conflict in previously peaceful countries are rare events and therefore hard to predict. To make progress in this hard problem, this project summarizes more than four million newspaper articles using a topic model. The topics are then fed into a random forest to predict conflict risk, which is then integrated into a simple static framework in which a decision-maker decides on the optimal number of interventions to minimize the total cost of conflict and intervention. According to the stylized model, cost savings compared to not intervening pre-conflict are over USD 1 trillion even with relatively ineffective interventions, and USD 13 trillion with effective interventions
The main advantage of the method is that it allows for a full ranking of risks for 190 countries worldwide within days of the release of news and without reliance on any data except for the UCDP data on armed violence. This circumvents problems of data availability in which information on institutions and economic activity are published only with a significant delay.
The free access published version is available here. An older version of the paper was published in 2019 as a CEPR Discussion Paper (13748) here. The framework has been used in a policy report for the Universal Rights Group to predict human rights atrocities. For news coverage by the CSIC see here. We have implemented an actual forecast of political violence based on the method for the Banco de España and the FCDO (UK). Thanks to this and funding from the CSIC we can publish monthly updates of outbreak risk on the webpage conflictforecast.org.