In many Voting Advice Applications (VAAs) a supplementary question concerning the voting intention of a VAA user is included. The data that are collected through this question can serve a variety of purposes; election forecast being one of them. However, it appears that the majority of VAA users who answer this question select safe choices such as “I prefer not to say” and “I am undecided”. In this study we investigate: (i) at what degree we can predict, with the aid of machine learning techniques, the voting intention of the above-mentioned users using as input their choices in the VAA policy statements, and (ii) whether there is any difference between the population of users who select “I prefer not to say” and those who select “I am undecided”.
The results of our study show that we can predict voting intention with accuracy higher than 60%, supposed that sufficient training examples for each party that participates in the elections exist so as to model each party users. Also, it appears that there is significant difference on the distribution per party for the users who select “I prefer not to say” and those who select “I am undecided”. As a consequence of these findings one would suggest that for effective election forecast it is required to (a) distribute the VAA users who select the previously mentioned choices in the voting intention question in a more sophisticated and intelligent way than that followed in traditional poll methods, and (b) the VAA users who select each one of those choices should handled separately.