Taking time seriously: Predicting conflict fatalities using temporal fusion transformers

Previous conflict forecasting efforts identified three areas for improvement: the importance of spatiotemporal dependencies and nonlinearities and the further exploitation of latent information in conflict variables, a lack of interpretability in return for high accuracy of complex algorithms, and the need to quantify prediction uncertainty. We predict conflict fatalities with temporal fusion transformers which have several desirable features for forecasting, addressing all these points. First, they can produce multi-horizon forecasts and probabilistic predictions, offering a flexible and nonparametric approach. Second, they can incorporate time-invariant covariates, known future inputs, and other exogenous time series which allows to identify globally important variables, persistent temporal patterns, and significant events. Third, this approach puts a strong focus on interpretability so that we can investigate temporal dynamics more thoroughly. Our approach outperforms benchmarks from an award-winning early warning system over several metrics and test windows and is a valuable addition to the forecaster's toolkit.