Predicting the costs of war

Abstract

The expected cost of war is a foundational concept in the study of international conflict. However, the field currently lacks a measure of the expected costs of war, and thereby any measure of the bargaining range. In this paper, I develop a proxy for the expected costs of war by focusing on one aspect of war costs – battle deaths. I train a variety of machine learning algorithms on battle deaths for all countries participating in fatal military disputes and interstate wars between 1816 and 2007 in order to maximize out-of-sample predictive performance. The best performing model (random forest) improves performance over that of a null model by 25% and a linear model with all predictors by 9%. I apply the random forest to all interstate dyads in the Correlates of War dataverse from 1816 to 2007 in order to produce an estimate of the expected costs of war for all existing country pairs in the international system. The resulting measure, which I refer to as Dispute Casualty Expectations, can be used to fully explore the implications of the bargaining model of war, as well as allow applied researchers to develop and test new theories in the study of international relations.