Forecasting change in conflict fatalities with dynamic elastic net

Abstract

This article illustrates an approach to forecasting change in conflict fatalities designed to address the complexity of the drivers and processes of armed conflicts. The design of this approach is based on two main choices. First, to account for the specificity of conflict drivers and processes over time and space, we model conflicts in each individual country separately. Second, we draw on an adaptive model—Dynamic Elastic Net, DynENet—which is able to efficiently select relevant predictors among a large set of covariates. We include over 700 variables in our models, adding event data on top of the data features provided by the convenors of the forecasting competition. We show that our approach is suitable and computationally efficient enough to address the complexity of conflict dynamics. Moreover, the adaptive nature of our model brings a significant added value. Because for each country our model only selects the variables that are relevant to predict conflict intensity, the retained predictors can be analyzed to describe the dynamic configuration of conflict drivers both across countries and within countries over time. Countries can then be clustered to observe the emergence of broader patterns related to correlates of conflict. In this sense, our approach produces interpretable forecasts, addressing one key limitation of contemporary approaches to forecasting.

 

Este artículo ilustra un enfoque para predecir cambios relativos a las víctimas mortales de los conflictos que ha sido diseñado para abordar la complejidad de los impulsores y de los procesos de los conflictos armados. El diseño de este enfoque se basa en dos opciones principales. En primer lugar, para tener en cuenta la especificidad de los impulsores y de los procesos de los conflictos a lo largo del tiempo y del espacio, modelamos los conflictos de cada país por separado. En segundo lugar, nos basamos en un modelo adaptativo, el Dynamic Elastic Net (DynENet), que es capaz de seleccionar eficazmente los predictores relevantes entre un gran conjunto de covariables. Incluimos más de 700 variables en nuestros modelos, a los que añadimos datos de eventos además de las características de los datos proporcionados por los convocantes del concurso de previsión. Demostramos que nuestro enfoque es adecuado y eficiente desde el punto de vista computacional para abordar la complejidad de la dinámica de los conflictos. Además, el carácter adaptativo de nuestro modelo aporta un importante valor añadido. Puesto que para cada país nuestro modelo solo selecciona las variables que son relevantes para predecir la intensidad del conflicto, los predictores retenidos pueden analizarse para describir la configuración dinámica de los impulsores del conflicto tanto entre países como dentro de los mismos a lo largo del tiempo. Los países pueden agruparse para observar la aparición de patrones más amplios relacionados con las correlaciones de los conflictos. En este sentido, nuestro enfoque produce previsiones interpretables, al abordar una limitación clave de los enfoques de previsión contemporáneos.

 

Face à la complexité des facteurs et mécanismes inhérents aux conflits armés, cet article propose une nouvelle approche de la prévision des changements en matière de conflits et des décès associés, basée sur deux grands principes. Tout d’abord, afin de prendre en compte la spécificité des facteurs et mécanismes des conflits dans le temps et l’espace, nous modélisons les conflits pour chaque pays, séparément. Ensuite, nous nous appuyons sur un modèle adaptatif (Dynamic Elastic Net, ou DynENet), capable de sélectionner avec précision les prédicteurs appropriés parmi un large éventail de covariables. Nos modèles comprennent plus de 700 variables et incluent, outre les données fournies par les organisateurs du concours sur les prévisions, des données sur les événements. Nous démontrons que notre approche est adaptée et suffisamment efficace sur le plan statistique pour faire face à la complexité des dynamiques à l’origine des conflits. Par ailleurs, le caractère adaptatif de notre modèle apporte une réelle valeur ajoutée. En effet, pour chaque pays, il sélectionne uniquement les variables capables de prédire l’intensité des conflits. Les prédicteurs ainsi retenus peuvent ensuite être analysés pour décrire la configuration dynamique des facteurs, que ce soit entre les pays ou au sein d’un même pays de manière diachronique. Il est alors possible d’effectuer des regroupements de pays, afin d’observer l’émergence de tendances plus globales reflétant les corrélations associées aux conflits. Notre approche permet donc d'obtenir des prévisions interprétables, surmontant ainsi l’un des principaux obstacles rencontrés par les méthodes actuelles.

Keywords:

• Armed conflict
• forecasting
• machine learning

Acknowledgments

We are grateful to Francesco Bonaccorso for excellent research assistance, and to two anonymous reviewers for their comments that greatly improved the manuscript. We would also like to thank Mike Colaresi, Håvard Hegre, and Paola Vesco for organizing the Violence Early Warning System prediction competition and workshop, and all competition and workshop participants.

Notes

1 See Web Page at: https://www.gdeltproject.org. GDELT “monitors the world's broadcast, print, and web news from different countries in over 100 languages and identifies the people, locations, organizations, counts, themes, sources, emotions, counts, quotes, images and events driving society daily”.

2 CAMEO scheme http://data.gdeltproject.org/documentation/CAMEO.Manual.1.1b3.pdf.

3 Namely: statement; appeal; express cooperation; consult; diplomatic cooperation; material cooperation; aid; yield; investigate; demand; disapprove; reject; threaten; protest; show force; reduce relations; coerce; assault; fight; use unconventional violence.

4 See GDELT codebook http://data.gdeltproject.org/documentation/GDELT-Event_Codebook-V2.0.pdf.

5 To produce the forecasts presented in this paper, we used these values for the QuadClass weight: 1 = Verbal Cooperation, 2 = Material Cooperation, 3 = Verbal Conflict, 4 = Material Conflict. One anonymous reviewer pointed out that this does not accurately follow an ideal ordering from less to more ‘conflictual’. This may be a subtle point, but it is certainly a correct one. Probably a better ordering could be obtained by trading weights 1 and 2, so that the scale would go from least (material cooperation) to most conflictual (material conflict). Or, we could have just used QuadClass 3 and 4 (conflict related) as possible weights. Because we designed the forecasting approach described in this article as part of the forecasting competition, we cannot revise ex post any potential flaws of the approach, which we plan to address in future implementations. Even though not an entirely accurate measurement choice, however, within the framework of a data-driven forecasting model this does not represent a fundamental bias, in our view (as it would instead be in the case of an explanatory model). An inaccuracy or error in the construction of a single indicator will reflect in the performance of the model, but not necessarily affect the reliability of the overall approach. Forecasting models are primarily evaluated by their performance (Toshkov 2016, 33). More specifically for what concerns our particular approach, as we mentioned above and further explain below, our algorithm only selects single (weighed) variables if they have some predictive relevance to the particular forecast that is being produced. In practice, this means that if one particular (weighed) variable has no predictive relevance (or it correlates to a similar variable with stronger predictive strength) it is dropped from the model (in favour of the variable with stronger predictive relevance). If GDELT variables weighed byQuadClass are selected in the model, that indicates that they are effective predictors of conflict in the particular context where the model retains them (although perhaps not as effective as they could have been, had they been more accurately designed).

6 Alternative sources of event data exist and in academic research, compared to other public or private sectors, they are more frequently employed than GDELT. We selected GDELT for our model because we wanted to design a forecasting system employing (at least some) data that would be updated in near real time, so as to feed a potentially live forecasting system also including real time alert functionalities. Other options would include the UTD Event Data, which were recently broadened to include near real time data (Kim et al., 2019; Solaimani et al. 2016).