# Mint Election Metrics projection: 289 for NDA, 118 for UPA

Our projections for the final tally based on leads treats early rounds of counting as virtual opinion polls. Now, this is not strictly correct, since the votes counted in one round do not strictly constitute a random sample of the population of the constituency—check out an earlier Election Metrics article on this topic.

However, in the absence of any other data, we can do worse than using the early leads in order to project the final voter count. The only thing we need to keep in mind is that the margin of error in early leads can be massive, and we should treat the data accordingly.

Rajeeva Karandikar of Chennai Mathematical Institute, who does the votes to seats conversion for CNN-IBN–CSDS has written an insightful blogpost on how to convert vote shares from opinion polls into seats. The most popular model for this is what is called as the “swing” model, where the country is divided into regions, and each such region is assumed to have a constant “swing” in favour of or against a particular party (this old Election Metrics piece shows this is not strictly correct, though)

In our “intraday” predictions on the day of counting we use a similar swing model.

The input into the model is the list of leads in a particular region (we have used states as regions for our analysis—again this is not strictly correct). We then simulate (using a Monte Carlo simulation to avoid messy mathematics) what the most likely set of swings for each party in order to produce the leads that we have seen thus far. Our model assumes a high degree of volatility in the leads in the early stages, and then we slowly decrease the volatility as more votes get counted (this is similar to the Machine Learning technique called Simulated Annealing.

Once we have determined this “ideal” set of swings for each region, we use these swings to project the vote shares of different parties in the region, and an aggregate of that gives our projections. In order to not project based on insufficient data, we do our projections only for regions where at least five leads have come in.

Based on data available at 9:30am, we projected a total of 289 seats for the National Democratic Alliance (NDA) and 118 for the United Progressive Alliance (UPA). Since the model has many sources of error, it may not be correct at the state level, so we are not publishing those results.