The election is over and done, and we have looked at our system’s predicted outcome vs the actual results. This was the first election that we’ve tried to predict that wasn’t a mainly two horse race, and also the smallest Twitter sample we’ve had by far both in volume and % of population represented (vs. the US, UK and French elections).
The results are (in % vote share):
Comparing our outcomes to the polls as of the time we did our calculations (late Thursday 21/9), honours were about even (see sum of cumulative % errors, bottom of 2 right hand columns). We think we caught the 2 main “surprise” trends better (Rise of AfD and fall of CDU/CSU), but we were worse on the rise of the FDP. (Note that the Polls kept on shifting throught Friday, Saturday and even Sunday as the AfD rise was becoming more clear to pollsters, our system had retty much held firm on those results for a good few days)
Our major error was the SPD support. Our system kept showing it as higher, we thought it was because it had fallen a lot since the height of its popularity mid-year and that our algorithms had to much smoothing and history, but in checking into the data it was still coming up as popular in the last 4 weeks and even just on the Thursday when we were doing our analysis. We did suspect that German Twitter, being a relatively small % of the German electorate, may have trended more liberal than average (that was the case in the US and UK in Twitter’s earlier days) so that may be a part of the cause but we are still not clear on the whole cause.
Big picture is that Germany fits the bigger European pattern we are seeing – the “mild left” dropping, going more left or more right, and a rise in the far right and far left. Typically the “mild right” has shifted more right to attempt to keep its voters who are shifting further right (eg. UK, Holland), the CDU seemed to not do so, and so it’s support fell. (Macron is a “one off” as he garnered a lot of “Non Le Pen” voters, but many are not natural bedfellows with each other. His problem will be to keep them all onside)
The other thing we need to think about is the AfD’s real level of support – huge numbers of tweets (c 50% of all tweets over the entire period) but with very low influence. We didn’t see a lot of bots (and our system scrubs most of them anyway), and neither did other watchers, so we hypothesize that this is in reality a motivated but unreported “grass roots” effect (“Influence” is heavily influenced by major media sources picking it up). What this implies is that underlying AfD support is still probably greater than the election gives them, so we’d hold to that underlying 15% as a more accurate picture, so still with with potential to grow if its gets a bigger voice now it is in power.
That is, of course, if it doesn’t fall apart in the meantime. Protest movements of all stripes have a long history of doing that though, as so brilliantly satirised by Monty Python’s Judean People’s Front sketch