This article proposes a dynamic decision model for situations in which a policymaker is facing a recurring risk and has the choice between early action (prevention) and late action (de-escalation). Dynamics, damages, and actions are modeled in a Markov model. This allows us to first derive three simple theoretical results. First, optimal prevention tends to crowd in de-escalation. Second, optimal de-escalation can crowd out prevention. Third, forecasts can make preventative efforts more likely. The framework is then applied to a global monthly dataset of armed conflict and risk forecasts over the period 2010-2022. Under reasonable assumptions on policy costs and effectiveness, the model reveals preventative policies that are currently receiving little systematic treatment by governments and international organizations but could be identified with systematic forecasts. The long-term benefit of such a policy shift would be significant.