I study a dynamic game of two-party competition in which party preferences are private information, exhibit serial correlation, and change with higher probability following defeat in elections. Assuming partisans care sufficiently about office, extreme policies are pursued with positive probability by the government when (a) both parties have a reputation for being extreme that exceeds a fixed level, and (b) elections are close in that both parties have similar reputations. Two qualitatively different equilibrium dynamics are possible depending on the speed with which the latent preferences of parties in government shift between moderation and extremism relative to the opposition. One dynamic produces regular government turnover and extreme policies along the path of play, whereas the other involves a strong incumbency advantage and moderate policies.