Taking time seriously: Predicting conflict fatalities using temporal fusion transformers

Abstract:

Previous conflict forecasting efforts identified two areas for improvement: the importance of spatiotemporal dependencies and nonlinearities and the need to exploit the latent information contained in conflict variables further, and that complex algorithms achieve high accuracy at the expense of interpretability whereas we should aim for more interpretability. Our approach predicts future fatalities with a novel transformer-based deep learning approach which tackles both the above points. Temporal fusion transformer models have several desirable features for conflict forecasting. First, they can produce multi-horizon forecasts and probabilistic predictions through quantile regression. This offers a flexible and non-parametric approach to estimate prediction uncertainty. 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 for our prediction problem. This mechanism makes them suitable to model both long-term and short-term dependencies. Third, this approach puts a strong focus on interpretability such that we can investigate temporal dynamics more thoroughly via temporal self-attention decoders.

Authors:

Julian Walterskirchen, Sonja Häffner, Christian Oswald, and Marco Binetti