Forecasting Monthly Fatalities via a Negative Binomial Distribution and Comparison with a Hurdle Model and Neural Networks

Abstract:

We choose a data-driven approach for our contribution to the 2023/24 VIEWS prediction competition. We focus on the country-month (cm) level and base our analysis on the data provided by the VIEWS team. Since it was communicated as the main metric in advance, our primary objective is to achieve optimal predictive performance with respect to the Continuous Ranked Probability Score (CRPS). We compare three modeling approaches that differ in their levels of complexity: a negative binomial distribution (NB), a hurdle model and feed-forward neural networks (NNs). Our model determination is based on test data that was available at the start of the challenge, i.e. the years 2018 to 2021.

Despite being the simplest of the three models, we find that the NB outperforms the other two more involved approaches in terms of the average CRPS in these years, while also being competitive or superior in the challenge’s secondary metrics. We therefore submit forecasts generated by a NB. More specifically, we use its 0.1, …, 99.9%-quantiles as our predictive samples. The resulting model is simple, straightforward, transparent and easy to interpret. It naturally models the conflict trap characteristic, yet by construction it is unable to predict the outbreak of conflicts or identify trends that have not occurred in the past. Regarding the additional test data for 2022 and 2023 that was released closer to the submission deadline, we find the NNs to be superior to our simpler approaches, which shows that they provide a promising path for potential future extensions of our work.

Authors:

Tobias Bodentien and Lotta Rüter