Citizen forecasts are increasingly used to predict elections. This increase is justified because citizen forecasting is among the most accurate election prediction methods, and its accuracy extends to subnational units, even when based on small and hence unrepresentative samples. The standard method for combining citizen forecasts is majority voting, which is optimal (minimum possible classification error) when accuracies are equal across citizens. Recently, weighted majority voting has been proposed for two-party systems, which is optimal when accuracies vary across citizens, but are equal across parties for each citizen. I introduce a probabilistic framework from pattern recognition that generalises weighted majority voting from two- to multi-party systems. This framework also includes two additional combiners, and I extend it by deriving another two. In total six combiners are derived subsequently from one another, by progressively relaxing and then eliminating assumptions about citizen accuracy or party accuracy. The framework also highlights that fewer assumptions come at the cost of more parameters to be estimated, which may make non-optimal combiners perform better than the optimal combiner. I apply the framework to citizen forecasting in the British multi-party system. Results from eight general elections slightly favour weighted majority voting, especially for constituencies with few citizens.