One of the good data sources I could find on google is this http://www.globalelectionsdatabase.com/index.php/index

This site contains election tables from countries across the globe. Tweets seem to be far fetched for Sentiment analysis, as you would have to cater to different politicians feeds to be able to map them for results.

On the other hand, electoral result prediction could be used with simple Regression techniques (Linear, Random Forest etc.). I was involved in a project related to demand forecasting and the problem seems to be somewhat similar. I was able to use various forecasting models for the same very conveniently, given that I had time-series data with me.

Hope this helps. Good Luck.

Thanks a lot Kashyap Barua

1. This problem will come under regression (supervised learning) 2. The problem is very complex because factors affecting the poll results are huge. A. Local factors -

• This will have local majority of people caste or religion as one variable
• And Candidates who participated and won there category based on caste.(Because in India this factors means a lot)
• Party previous performance or local Leader's hold in that area.

B. Global factors

• Party at national level popularity
• National level leader
• Money or funding

So how will you generate data.

• Popularity at local and national level u can take from surveys all most all news channel will have this data.
• Big face and their popularity you can get in same way
• Caste or religion based data you can collect from census data.

Thanks a lot Venkatesh Gauri Shankar .

Regarding your question, I recall an interview about how people choose their president. A man said "this guy looked handsome". Some people may say "this guy fits into my taste".

Above all, I mean that in addition to some obvious data sources like local factors and global factors as mentioned by Venkatesh Gauri Shankar , you may need to consider 1. the look of the candidate 2. personality of the candidate and so on.

Sometimes such dimensions of data that seems irrelevant may give us surprising results.

In addition, you need to care about the inexplicit bias in your data. The data sources from like survey may show bias and results in wrong results.

Dear Yuanrui Fan,

Thanks for suggesting the personality factors of candidates.

Agent based modelling (ABM) is an option. You program the behavior of N kind of electors and free them in a field to see how they build a social behavior (see simulations in NetLogo).

Also clustering is a nice technique if you have enough data about a diverse population.

Of course if you have target variables, neural networks are a good option.

Thanks a lot Jose Machicao .