In this paper, we assess the predictive quality of different dynamic Bayesian forecasting models suited for modeling the compositional data that results from trial heat polls in multi-party systems. In particular, we compare different spec- ifications of dynamic linear models (DLM) fitted on log-ratio transformed data. In addition to existing random walk models [Jackman, 2005, Linzer, 2013, Walther, 2015], the DLM framework allows for model specifications with local trends and covariance structures that reflect the correlated evolu- tion of party support over time. We assess the point predictions of different DLM specifications and the coverage of the corresponding predictive inter- vals in the run-ups to the German Federal Election of 2013. Comparing polls the DLM based forecasts, we find that the latter generate predictive inter- vals that more accurately represent the uncertainty about the final outcome far ahead of election day. Only within the last month prior to election day, polls offer reliable information about the expected election outcome. With these findings, we provide guidance for model selection when modeling dy- namic compositional data within or beyond the specific context of elections forecasting.