This paper studies the performance of alternative prediction models for conflict. The analysis contrasts the performance of conventional approaches based on predicted probabilities generated by binary response regressions and random forests with two unconventional classification algorithms. The unconventional algorithms are calibrated specifically to minimize a prediction loss function penalizing Type 1 and Type 2 errors: (1) an algorithm that selects linear combinations of correlates of conflict to minimize the prediction loss function, and (2) an algorithm that chooses a set of thresholds for the same variables, together with the number of breaches of thresholds that constitute a prediction of conflict, that minimize the prediction loss function. The paper evaluates the predictive power of these approaches in a set of conflict and non-conflict episodes constructed from a large country-year panel of developing countries since 1977, and finds substantial differences in the in-sample and out-of-sample predictive performance of these alternative algorithms. The threshold classifier has the best overall predictive performance, and moreover has advantages in simplicity and transparency that make it well suited for policy-making purposes. The paper explores the implications of these findings for the World Bank's classification of fragile and conflict-affected states
