We provide a statistical model to predict and understand the number of fatalities caused by political violence in the administrative regions in Africa one quarter ahead in time. We address several modeling challenges like zero-inflation, misclassification of the conditional variance, by using a Hurdle model based on a negative binomial regression. By fitting all distribution parameters of the zero-truncated negative binomial distribution, we are able to tackle possible over-dispersion and give measures of uncertainty regarding the predictions and construct prediction intervals. By using model-based boosting, which is the same as functional gradient descent, a sparse model is obtained. The spatial and temporal structure is taken into account as well as random effects to also consider country-specific effects. We compare the model using only linear base-learners with more complex structures, like penalized spatial Splines and random slop es, by computing the out-of-sample log-likelihood and other evaluation criteria. The models are evaluated in the year 2019 and the first two quarters of 2020 separately, because of the global pandemic, showing the model robustness in these two circumstances.     «

We provide a statistical model to predict and understand the number of fatalities caused by political violence in the administrative regions in Africa one quarter ahead in time. We address several modeling challenges like zero-inflation, misclassification of the conditional variance, by using a Hurdle model based on a negative binomial regression. By fitting all distribution parameters of the zero-truncated negative binomial distribution, we are able to tackle possible over-dispersion and give m...     »