Probabilistic Forecasting of Conflict Fatalities: Historical Quantiles vs. Bayesian Penalized Regression

Abstract:

In the following, we present the general model selection and tuning process for our participation in the 2023/24 VIEWS prediction competition (Hegre et al., XXX), where we submit probabilistic predictions for conflict fatalities on the country-month (cm) level. As a first step, we narrow the scope of potential modeling approaches by considering two competing model classes, which differ considerably in their general level of complexity. The first model (class) constitutes a rather simple and transparent baseline, which we expect to produce solidly performing and stable predictions. The second model class is based on Bayesian structured additive regression models, which theoretically have the capability to outperform the baseline due to the higher available degree of flexibility to model
conflict fatalities.

After training and tuning both model classes via several specifications on the training set, we then choose the singular model which attains the lowest CRPS score on the years 2018-2021. Concretely, we therefore give preference to the model with superior performance in terms of the main metric by which submissions will be scored. This also means opting for a more data-driven model selection process and, in particular, giving preference to the baseline model, should the more sophisticated Bayesian model not demonstrate superior performance.

Throughout the following, we denote the outcome variable by yi,t, where i refers to the country and t to the month. In a concrete prediction context, we additionally denote the forecast horizon by s, where yi,t+s is the s-step-ahead quantity to be predicted. We rely exclusively on VIEWS-provided data to train models and produce predictions. Additionally to the centrally provided data, some specifications of our Bayesian model utilize neighborhood data, which was kindly also provided by the VIEWS team.

Sections 2 and 3, respectively, introduce the baseline and Bayesian model in more detail and Section 4 summarizes our results. Section 5 concludes.

Authors:

Simon Drauz and Friederike Becker