Dynamic policy games feature a wide range of equilibria. This paper provides a methodology for obtaining robust predictions. We focus on a model of sovereign debt, although our methodology applies to other settings, such as models of monetary policy or capital taxation. Our main result is a characterization of distributions over outcomes that are consistent with a subgame perfect equilibrium conditional on the observed history. We illustrate our main result by computing, conditional on an observed history, bounds across all equilibria on: the maximum probability of a crisis, means, variances, and covariances over debt prices.