This research aims to establish potential opportunities for policy intervention by asking the following question: if we could predict conflict 5-years out, what would separate the predictable failures from the unexpected successes (i.e. places where conflict was expected but did not happen) and what would separate the predictable successes from the unexpected failures (i.e. places where conflict was not expected that experience it). The idea is to move beyond standard conflict prediction variables to identify previously-unrecognized opportunities for preventive action within a time-frame sufficiently long for significant action by the UN, World Bank, and other international organizations. Put differently, our goal is to provide policy recommendations by examining what led to unexpected resilience to conflict in some countries and what happened in places with low risk based on observable characteristics. Existing studies focus in on a set of states which did become affected by violent conflict without comparing them to a set of peers based on objective and replicable criteria. This study addresses this issue by looking at the set of developing countries at the end of the Cold War, predicting a priori which ones would be most likely to become affected by violent conflict, and identifying surprising successes and surprising failures. Essentially, we use machine learning to approximate what a country team would do in terms of predicting conflict risk five years out. This approach allows us to compare states which did become affected by violent conflicts (or remained in violent conflict) to states which did not (or that emerged from violent conflict). On the basis of careful comparison between matched cases, we highlight systemic differences and assess potential policies that can reduce the risk of conflict.