This paper tests the feasibility of local-level violence forecasting. We apply standard prediction models to new data from 242 Liberian communities to investigate whether it is to possible to predict outbreaks of local violence with sensitivity and accuracy, even with limited data. We first trained our models to predict communal, extrajudicial, and criminal violence in 2010 using 2008 risk factors. We then made forecasts of violence in 2012, before collecting data. Our model predicted up to 88% of actual 2012 violence. This came at the cost of many false positives, for overall accuracy of 33 to 50%. Policy-wise, states and peacekeepers could use such predictions to prevent and respond to violence. The models also generated new stylized facts for theory to explain. In this case, ethnic cleavages and power-sharing predicted violence, while economic variables typically did not. We illustrate how forecasting can be widely more applied to micro-level conflict data.