Can word-of-mouth predict the General Election result?

The forthcoming General Election is probably going to be one of the closest in recent UK history, with the pollsters suggesting various hung parliament and minority Labour or minority Conservative government scenarios. During the General Election campaign, Tweetminster is conducting an experiment around whether activity on Twitter correlates to electoral success.

Our inspiration for this experiment comes from last year’s General Election in Japan, when a group of software engineers and PhD graduates from Tokyo University undertook a study analysing the correlation between ‘online buzz’ and election results. The aim of the study was to assess if word-of-mouth mentions of candidates could help to predict which ones would be successful. The study found that in a majority of constituencies the most mentioned candidate won the seat (see References below).

We thought it would be interesting to run a similar experiment to the Japanese study in the UK using Twitter. From now until the election we will be tracking the most mentioned (i.e. posts and conversations about) constituencies and candidates on Twitter and using this data we will try to map the correlation between buzz, word-of-mouth and the eventual election results through a predictive model.

Today, to kick-off the study, we’re publishing a starting set of findings and the methodology that we’ll be adopting.

This paper sets out the initial findings of our experimental model, which we will update as the campaign proceeds.  At this stage, our model suggests that the overall election result could see a small Labour majority or a hung Parliament, with the closely-fought contest between the Liberal Democrats and the Conservatives in a number of marginals in the South West tilting towards the Lib Dems; with Labour and the Liberal Democrats performing better in London than recent polls have shown; with declining SNP support in Scotland; and with the role of other parties in key seats all influencing factors in shaping our predictions.

Initial top-line predictions include an analysis of key target seats, a top-line party breakdown based on the most mentioned candidates in the 376 constituencies represented on Twitter - CONSERVATIVES 34%; LABOUR 35%; LIBERAL DEMOCRATS 22% OTHERS 9% - which, assuming margins of error, would encompass various hung parliament scenarios with Labour short of seats to a Labour majority of 14 seats.

Our data set is fed from all the constituencies represented on Twitter:

1.     Constituencies with a candidate using Twitter

2.     Frequently mentioned constituencies

3.     High profile constituencies, i.e. key marginals and cabinet/shadow cabinet members’ constituencies that are mentioned on Twitter.

The study will be a dynamic analysis - we will update the predictions and track the variations in predicted election results as polling day gets closer, allowing us to see if the passage of time affects any significant shifts in predicted outcomes.  For example: Will the Leader’s TV debates in mid-April make a difference to today’s predictions?

As discussion intensifies towards the election we expect to see the model reflect any changes in the balance of online buzz.

Two million tweets (and counting) are being processed and analysed for this study. The data set will be updated as new candidates join Twitter during the course of the campaign, findings and variations released throughout the campaign and the final report will be published after the election.

Please note that the scope of this exercise isn’t to compete with polling methodologies - it is an experimental study that aims to use predictive modelling around a dynamic data set to determine if there are correlations between word-of-mouth on social media and election results. All predictions are made on an experimental basis and the reliability of the method for predicting election results will be assessed once the study is completed.

The paper can be found here.

Update - our week two updates have been released. The paper has also been updated accordingly.

Posted at Tue, Mar 30th 2010, 09:00

blog comments powered by Disqus

Tweetminster blog