We emphasize some strengths of our approach. First, we employ out-of-sample forecasting by training the model parameters on genocide onset in a previous period and using it to forecast genocide onset in a subsequent period. This gives a rigorous test of ‘prediction’ in that it guards against over-fitting the model on the very events that are being forecast. Second, we use a global dataset with annual measures for all countries in the world. Third, for some forecasts we model a process in two stages, instability and genocide. This helps us account for the process often assumed to lead to mass killing, while allowing us to produce forecasts for all countries in the world in out-of-sample ‘future’ years, rather than restricting our ‘forecasts’ to countries with ongoing instability or a relatively small number of selected ‘control’ countries. Fourth, for some forecasts we use machine-learning based approaches to better account for the complexity of the process leading to genocide onset.