Scholarly efforts to forecast parliamentary elections have targeted the national level predominantly, disregarding the outcomes of constituency races. In doing so, they have frequently failed to account for systematic bias in the seats–votes curve, and been unable to provide candidates and campaign strategists with constituency-level information. On the other hand, existing accounts of constituency-level election forecasting suffer from data sparsity, leading to a lack of precision. This paper proposes a correction–combination procedure that allows for the correction of individual constituency-level forecast models for election-invariant bias, then combines these models based on their past performances. I demonstrate the use of this procedure through out-of-sample forecasts of 299 district races at the 2013 German federal election.