Accurate and interpretable prediction of vote results during ballot counting is, among others, relevant to political parties, polling agencies, news outlets, and citizens. Such predictions can help political parties and interest groups to enhance their campaigning efforts and polling agencies and news outlets to optimize their surveying efforts. In this talk, I will talk about SubSVD, an algorithm that can accurately predict national vote outcomes from partial regional outcomes that are revealed sequentially. SubSVD combines matrix factorization techniques and linear models to make predictions while being interpretable. The algorithm is deployed as part of predikon.ch and used to predict outcomes of Swiss referenda in real time. I will additionally discuss how Bayesian models can help avoid otherwise misleading predictions in the context of real time vote prediction.