Terrorism occurs in around 75 per cent of civil conflicts. Previous studies have explained variation in the number and occurrence of terrorist attacks by rebel group, region, and country per month, year and over longer periods of time. They identify group, country and regime characteristics that make terrorism more likely. However, these measures have not been tested for their predictive performance and theoretical value. This study will fill this gap by testing the predictive power of existing indicators used in the literature to explain the occurrence of terrorism and the number of terrorist attacks and victims per month by known rebel groups. Random forests are used in combination with interpretable machine learning techniques to illustrate findings intuitively and to go beyond standard measures of accuracy, precision and recall. The empirical evidence suggests that while some variables have very high predictive power, other previously identified explanatory factors have limited predictive power. In addition, interaction effects are discovered with implications for future research. In light of these results, data-driven theory building and refinement are a valuable addition to the research process beyond the study of terrorism in civil conflicts.