Since January 2014, Fabio Wasserfallen and I have been running surveys in Switzerland before each ballot measure for a large newspaper. We usually conduct three waves, about fifty, thirty, and ten days before the actual vote takes place. Unlike traditional polls we rely on online surveys and employ extensive statistical modeling to produce good estimates of public opinion.
The last survey is always reported 10 days before the final vote. One way to assess our estimates is to compare the estimated support ten days before the vote with actual outcome. We compare our results (Tamedia) with our major competitor’s estimates (GfS/SRF). Past performance is good and in 6 out of 10 times we are closer to the final outcome than the traditional phone survey. When we use the last two polls to create a linear extrapolation we outperform the other poll in two out of three times (comparison data).
We are also working on a prediction model (not published) which is based on our three waves and additional inputs. We currently have an average absolute out of sample error of 3.8 % points.
A short report with a more detailed description of what we do (and how this contrasts with traditional polling approaches) is available here – unfortunately, it is only available in German. Do not hesitate to get in touch with me if you have any questions about this polling work.