A Multimodel Ensemble for Forecasting Onsets of State-Sponsored Mass Killing

Abstract

This paper describes the development of a multimodel ensemble that is designed to estimate annually for every country worldwide with a population larger than half a million the risk of an onset of state-sponsored mass killing. Prior applications of statistical methods to the problem of early warning on mass atrocities have sought to develop a single best model for forecasting, and they usually have attempted to test hypotheses about the causes of mass atrocities as they go. By contrast, the research described in this paper focuses exclusively on the problem of forecasting, and the solution we have adopted involves an ensemble approach that averages predictions from several models. Over time, this ensemble should produce more accurate predictions than any single statistical model or expert forecaster would, and it lays a foundation for further learning by leaving the door open to incorporating probabilistic forecasts from new models and other sources as they are developed.

Keywords: genocide, atrocities, forecasting