We examine whether features highlighted as important for mediation in existing research allow us to predict when we will see mediation and likely success out-of-sample. We assess to what extent information about the characteristics of the conflicting dyads and conflict history can be evaluated ex ante and improve our ability to predict when conflicts will see mediation as well as when peaceful solutions are more likely to follow from mediation. We justify that the information used to generate predictions through the model can be assessed ex ante, using the ongoing conflict in Syria as an example. Our results suggest that a two-stage model of mediation and success seems to do relatively well overall in predicting when mediation is likely to occur, but notably less well in predicting the outcome of mediation. This may reflect how ex ante observable structural characteristics are likely to influence willingness to mediate, while the outcome of mediation to a large extent will be influenced by unobservable characteristics or private information and how these are influenced by mediation. We discuss the usefulness of out-of-sample evaluation in studying conflict management and suggest future directions for improving our ability to forecast mediation.