Interesting, See, everything that happens is a function of countless number of inputs ( features). While modelling to make predictions, one needs to take into considerations as many of these inputs/ features as possible.
My worry in this case is that you probably are discounting some really important features like sentiments of people, no. of recent rallies conducted by each party, alliances formed, etc. For some of these you can probably mine twitter and facebook data.
Although, not discouraging you'll have start with looking for what is the most relevant data you can find related to election results from the past.
Another approach might be to use ARIMA models if you only want to use the election results. Will they give good results. Not sure, but you can definitely try.