Using past violence and current news to predict changes in violence

Abstract

This article proposes a new method for predicting escalations and de-escalations of violence using a model which relies on conflict history and text features. The text features are generated from over 3.5 million newspaper articles using a so-called topic-model. We show that the combined model relies to a large extent on conflict dynamics, but that text is able to contribute meaningfully to the prediction of rare outbreaks of violence in previously peaceful countries. Given the very powerful dynamics of the conflict trap these cases are particularly important for prevention efforts.

Este artículo propone un nuevo método para la predicción de escaladas y desescaladas de violencia a través de la aplicación de un modelo basado en los antecedentes del conflicto y las características propias del texto. Las características del texto se generan a partir de más de 3,5 millones de artículos de periódicos mediante el uso de lo que se denomina “modelo de tópicos”. Demostramos que, si bien este modelo combinado hace referencia a una extensa dinámica del conflicto, el texto es una contribución relevante que permite predecir los estallidos de violencia inesperados en países que antes eran pacíficos. Dada la dinámica de gran intensidad característica de la trampa del conflicto, estos casos son de especial importancia en lo que se refiere a las iniciativas de prevención.

Dans cet article, nous proposons une nouvelle méthode destinée à anticiper les escalades et désescalades de violence grâce à un modèle reposant sur les antécédents conflictuels et sur des caractéristiques textuelles. Ces caractéristiques sont extraites à partir de plus de 3,5 millions d’articles de presse à l’aide d’un modèle thématique (topic model). Nous montrons que si ce modèle mixte s’appuie largement sur les dynamiques conflictuelles, les données textuelles peuvent être très utiles en vue d’anticiper les rares explosions de violence dans les pays habituellement pacifiques. Étant donné la puissante dynamique qui sous-tend les conflits récurrents, les exemples exposés revêtent une importance particulière dans une optique de prévention.

Keywords:

• Armed conflict
• civil war
• forecasting
• machine learning
• text analysis

Acknowledgment

We thank the organizers of the forecasting competition for their guidance when developing this project. This is an important public service. We also thank the editor and referees for their feedback to an earlier version of this article. All errors are ours. We thank Bruno Conte Leite and Luis Ignacio for excellent research assistance.

Data availability statement

The replication files to this article can be found in the Dataverse page https://doi.org/10.7910/DVN/BW7UV4.

Notes

1 For overviews see, Schrodt, Yonamine, and Bagozzi (2013), Ward et al. (2013), and Hegre et al. (2017).

2 For an interactive representation of an LDA estimated on a similar corpus see the webpage https://conflictforecast.org/.

3 We limited the grid search to this set of hyperparameters by searching across a wider range in a preliminary analysis. Due to the computational complexity and processing times, we did not do an overly exhaustive grid search for every round.

4 Only for 2020 we do not report the actual data for better visibility.

5 Mueller and Rauh (2018, 2022) analyse the role of text in predicting onsets in detail finding that topics associated with violence (like the military, terror or even peace agreements) directly increase before conflict breaks out. Topics that capture judicial procedures, diplomacy, economic activity or trade tend to decrease before violence breaks out.

Additional information

Funding

Hannes Mueller acknowledges financial support from the Banco de España and the Spanish Agencia Estatal de Investigación (AEI), through the Severo Ochoa Programme for Centres of Excellence in R&D Barcelona School of Economics CEX2019-000915-S). This work was also supported by Fundación BBVA; Spanish Ministry of Science and Innovation.