We build on previous quantitative analysis of genocide/politicide, mass killing, civil war, and political instability to inform our forecasting model. We develop a two-stage approach which considers genocide and politicide as emerging from already unstable political situations. We use variables found to be significantly associated with instability and/or genocide/politicide in previous studies, but we give preference to the forecasting power of each factor, rather than statistical significance. We also introduce a number of time-sensitive factors which to the best of our knowledge have not been examined previously, including political assassinations, election periods, changes in military force numbers, and changes in political institutions.
Our two-stage approach allows us to include data for all available countries and years, while still acknowledging the theoretical insights from the literature that genocide emerges in unstable situations. Thus we produce forecasts for all countries in the international system in a given year, but the likelihood of genocide is informed by the likelihood of instability. Our analyses using standard econometric approaches have employed both Probit selection models and also distinct Probit models for each stage, incorporating the estimate of the future probability of instability as a parameter in the model for future genocide/politicide onset. However, our machine-learning based approaches, which produced the forecasts in the reports and future forecasts section of this website, incorporate variables from stages 1 and 2 into a single equation, for example using a Generalized Additive Model or a Boosting approach. Thus we have found that an explicit two-stage approach does not improve forecasting performance with more nuanced machine-learning techniques, although it is still important to include factors associated with both political instability and genocide onset.