We applied evidence-based principles to forecast U.S. presidential elections. The primary strategy was combining. We mechanically combined forecasts that used different methods and different data to produce the PollyVote. This involved unweighted combining within and across four categories of methods (polls, a prediction market, quantitative models, and a survey of political experts). We applied the PollyVote prospectively to the 2004 and 20008 elections, and retrospectively for three elections from 1992 to 2000. Improved accuracy was achieved by combining within components as well as across components. The errors were especially low in the 2004 and 2008 elections, when the PollyVote drew upon more methods and more data; the average election-eve error was 0.4%. Across the whole forecasting horizon, the PollyVote correctly predicted the winner on 99% of the 957 days on which comparable forecasts were made; no other method came close to this accuracy. Our future work is focusing on making longer-term forecasts (e.g., before candidates are selected) and in developing methods that can be used to select candidates and to inform campaign strategies. Given the large number of variables for these problems, we use “index models.” I will report on initial results for our PollyBio and PollyIssues models.