Talk by Shah Shlok

The prediction of armed conflict is challenging but important because of its potential to facilitate early prevention and to focus mitigation efforts by policymakers. A fundamental challenge in conflict prediction is the often poor estimates of conflict fatalities, which are difficult to collect. Based on previous observations that the distribution of conflict fatalities is close to a power law, we propose a Bayesian maximum a posteriori estimation scheme for data reconstruction. To further reduce the impact of noisy and low-resolution data, we coarse-grain variables to appropriate levels of detail by taking into account the information gain across resolution scales. Then, to explore the predictability of the data, we fit random forest ensemble models to the coarse-grained predictors and fatality targets and find that model performance and decision boundaries do not usually change significantly beyond deduced levels of resolution. Finally, noting that conflicts are sequences of related events, we employ LSTM neural networks and discover non-trivial gains in performance.