1. Our first exercise was to database the election results of the previous four election cycles. This allowed us to perform a conditional logic analysis that predicted the odds of winning for each party and alliance within each constituency. As part of this analysis we also made predictions and updates to the data to ensure the delimitation of 2008 was taken into account. Ultimately, we had a database with seat safety for each constituency as well as the average winning margin for that party. The report based on this analysis can be seen in this blog post.
2. Now that we had the historic data captured, we focused on the pollster predictions coming through the media. We did not use any national predictions, but rather focused on the state predictions. As has been seen in previous predictions, we found that a number of leading pollsters had been relatively accurate in their by-state predictions. The best of these pollsters was the CSDS. Of course, 2004 had seen pollsters perform poorly, but even then many of their state predictions were fairly accurate. There were, obviously a few exceptions. Our blog post that goes into further detail on our analysis of the pollsters can be seen here. During this 2014 election cycle we databased the predictions made by CVoter, CSDS, Hansa, IPSOS and AC Nielsen. For each State and Union Territory we databased their seat predictions.
3. For each seat prediction, we used the historic data to determine what the poll margin win would be within each constituency or state. From these margins, we then determined the Median and SEM of winning the seat.
4. We finally created a dataset for each of the three alliances containing the Median and SEM for their winning margin within each of the 543 seats. We used this data to determine the exact number of winners, covering all ways of reaching that number given the win probability. We adapted the approach presented here.