Mainstream electoral forecasting usually fails to provide reliable predictive intervals. Even exit polls are typically presented as point estimates and pundits are left to proclaim their own assessment of the poll's accuracy. Conversely, a satisfactory forecast consists of a full predictive distribution over the outcome of the election. In that spirit, I discuss how to turn disparate data of varying quality into such a forecast. I will focus on the application of established statistical techniques, such as post-stratification and Bayesian hierarchical modelling.