A Zero-Inflated Poisson Generalized Additive Model for Forecasting Conflict Fatalities

Abstract:

Conflict data often face a seeming limitation: outbreaks of major episodes of political violence are, thankfully, rare. This means that the modal value of the dependent variables is often zero, but the tails of conflict distributions can also be strongly right skewed. This poses forecasting problems for both practitioners, who want to know whether any conflict might occur, and for methodologists, who are concerned with accurately predicting how much conflict there may be, as predictions may be strongly biased downwards by this zero-inflation. We propose a semi-parametric zero-inflated constrained generalized additive model for forecasting conflict. This model proceeds by first utilizing a binomial distribution to predict whether any conflict will be observed, then, given a positive prediction of any conflict, utilizes a Poisson distribution to predict how much conflict there may be. Our model achieves predictive accuracy as measured by the continuous ranked probability score comparable or better than the VIEWS benchmarks for nearly all years in the test set. We generate predictions for the VIEWS forecasting window of June 2024-July 2025 by simulating from the model’s predictive distribution.

Authors:

David Muchlinski and Chandler Thornhill