David M Rothschild on Posted on

The final voting for the India’s general election concludes at 6:30 PM local time on Monday, May 12 and, “as the voting closes in India, no one knows what the results will be.” There are total of 814 million eligible voters in this massive display of democracy, which is why voting is spread out over more than a month (not enough polling machine and election workers, so they need to move around the country). It will take an additional four days to count the final vote tallies, due to be released on Friday, May 16, although we expect exit polls to be released as the voting closes.

While readers from the United States are accustomed to intense and accurate coverage of the election polls through Election Day (the US is already inundated with 2016 election polls), India bans polling from being released during the voting portion of the campaign. Microsoft Research saw this as an opportunity to test a new crowd sourcing prediction game to see how well we could do at predicting the unpredictable.

The Indian Elections Game, joint work between Microsoft Research India, including Shipra Agrawal and Nikhil R. Devanur, and Microsoft Research New York City, including Miro Dudik and myself, was released just before voting began on April 7 and ran through today’s final voting day. Nearly 4,000 users created their predictions on the number of seats that any given party, or combination of parties, would get in any of the 35 voting states and union territories. For example, a user may look at Delhi’s seven seats and predict that BJP would get between 4 and 5 seats while Congress (INC) would get between 0 and 1 seats.

As the crowd makes predictions on the outcome of all 35 voting states, the national parties, and national alliances, the backend of the game examines the crowd’s predictions and creates the probability of any given outcome occurring. In this example, since game’s aggregation of the crowd’s predictions currently concludes there was just less than 50% likelihood of this combination occurring, a prediction that wagered 1 point on BJP getting 4 or 5 seats and Congress getting 0 or 1 seat would yield 2.28 points if correct and 0 points if incorrect.

Computing the probability was not trivial, as the game had to learn how the different parties interacted within a given district and between districts. And, unlike the United States where there are just two major parties, there is large number of viable parties, some national and some regional, competing for the 543 seats. The potential number of outcomes is huge! In future blog posts, over the next few days, we will go into more depth about how all of that was constructed.

But, right now, just as the election closes, and the exit polls are released, we wanted to share with you the final probable outcomes that our game generated from the approximately 4,000 users who played over the last month.

We expect the opposition party BJP to carry 228 of the 543 seats and their national alliance, the NDA, to carry 270 seats. The party currently in power, Congress is expected to carry just 87 seats and its alliance, the UPA just 110. The Third Front alliance is expected to carry 80 seats. That leaves 83 seats to smaller parties that are not formally aligned with any of the three major alliances. The national party predictions have a margin of error plus or minus 7% of seats and the national alliances plus or minus 10% of seats.

This prediction is generated by considering each voting states independently and then considering how they aggregate up together. Below we list each of the 35 states and our individual predictions:

Please check back over the next few days we discuss this game and its predictions in more detail!