The research paper that explains the combining approach used in the PollyVote has been updated. The latest version can be downloaded for free from the SSRN website.

Combining Forecasts: An Application to Elections

Graefe, A.; Armstrong, J. S.; Jones, R. J. Jr. & Cuzán, A. G.

Abstract. Combining forecasts is simple, inexpensive, and effective, yet few organizations use it. This occurs because the benefits of combining are counter-intuitive and because people are unaware of the empirical evidence. Here we summarize the literature on the effectiveness of combining by assessing the conditions under which it is most valuable. Using data on U.S. Presidential elections from 1992 through 2008, we then examine the error reduction obtained by averaging forecasts within and across four groups of methods (polls, the Iowa Electronic Markets, quantitative models, and expert judgment). The gains in accuracy from combining increased with the number of component forecasts used, especially when these forecasts were based on different methods and different data. Combining yielded error reductions ranging from 10% to 61% compared to the average errors of the uncombined individual forecasts; this improvement is substantially greater than the 12% that had been previously reported for combining. We also show combining is especially effective in situations involving high uncertainty.