Predicting violent conflicts atdifferent time horizons : Examining model weighting strategies

Sammanfattning

In this thesis the effectiveness of four different ensembling strategies in conflict forecasting iscompared: Using no ensembling, using equally weighted models, using a genetic algorithm tooptimize the weights, and using sequential least squares programming (SLSQP) to optimize theweights. As a first step, 36 Random Forest Classifiers, specialized on 36 different time horizonsare trained, to detect different states of conflict: peace, escalation, deescalation and war givenindependent variables regarding, among others, conflict history, political stability and economics.As the next step, predictions with the four different ensembling strategies are performed on a threeyear period unseen by the models. As a last step, two Random Forest Regressors are trained topredict fatalities resulting from the conflicts. Overall, with the setup and data partitioning of thisproject combining models of different time horizons provided only limited benefits to the accuracy ofthe ensembles with regards to predicting conflict state and even decreased accuracy with regardsto predicting fatalities. Neither of the optimized strategies provided any better results than usingequally weighted models.