Keywords: armed conflict, machine learning, prevention, dynamic optimization, information gain
JEL codes: O1, F5, F1, L8
Abstract
We study the gains of improving forecasting when a policymaker is facing a recurring risk and has the choice between a preventive early action and a de-escalating late action. We first introduce a simple two-stage Markov model to illustrate how prevention and de-escalation interact. We then study the role of forecasting for optimal armed conflict prevention in a 12-stage model which is calibrated using a large cross-country panel. Prevention benefits are substantial but critically depend on the systematic use of forecasting. The information rent of using a forecast is larger than 60% of GDP. In line with the theory we find that de-escalation policies reduce the incentives for prevention, whereas prevention increases incentives for de-escalation.