How to predict the outcome of elections in a multi-party setup?



I want to create a model to predict election outcomes in a multi-party elections. I want to do so at each constituency. We have the following variables:

  • Outcomes in the constituency for last 5 elections
  • Demographics of the candidates who contested the elections from these constituency
  • Social media sentiments through twitter
  • Relative presence of the candidates on the social media (vs. other candidates from same constituency)
  • Share of Voice and some indicators on the spend in the run up to elections
  • Average demographics of people in the constituency (e.g. Age, Income, family size)

We want to predict the percent votes received by each of these parties. Since this is not a bi-party system, we can not use Logistic regression.

What should be the right modeling technique to make these predictions?



Multi-class prediction models are common. You can use something like a SVM or Random Forest. However, I see a bigger concerns here : the number of datapoints. You simple have 5 data points for every constituency on which I don’t know how will you manage a predictive model. But, in case you combine all constituency, you stand a chance. In this case the primary assumption is that all the constituency are similar. Other way is to combine like constituency and then build multiple models.

Hope this helps.