SPPU students developing new alternative exit poll model
Three students of the Statistics Department from the Savitribai Phule Pune University (SPPU) have used the ‘random forest model’ to experiment on the prediction of exit polls through the data mining process.
Pune: Three students of the Statistics Department from the Savitribai Phule Pune University (SPPU) have used the ‘random forest model’ to experiment on the prediction of exit polls through the data mining process.
Vinay Tiwari, R Viswanath and Sharad Kolse, students of the second year MSc (Statistics) who share a common interest in politics, decided to conduct a study of methods used during exit polls. They have applied the new method to gauge the outcome of the Lok Sabha elections 2019. They said this model is used for the first time to predict election results at the academic level.
Data-mining for better accuracy
“In the Western countries, poll predictions are more accurate than ours because they use machine learning / artificial intelligence to predict results. If these technologies are incorporated in the prediction of election results in India, it will provide more accuracy,” said Tiwari.
The students have been working on this project for the past six months and have managed to gain data from 10 states, including Maharashtra. They have predicted that in Maharashtra, the BJP may get 17 to 23 seats; the Shiv Sena could bag 16 to 21 seats; while the NCP may win 3 to 9 seats and the Congress 1 to 6 seats.
The study is still underway. However, inferences derived till date have been encouraging them to state that the model could be applied to predict exit poll results in the future. “Until 2014, an overall prediction of the state exit polls was provided. Through this method, we can obtain region-wise details,” said Tiwari.
The random forest model is a classification method through the combination of learning models that increases the overall result. It is used for commercial purposes to conduct consumer surveys.
The students obtained the data about the elections such as number of votes, number of seats, etc, from Lokniti, a research programme by the Centre for the Study of Developing Societies (CSDS). “We availed most of the data from the surveys conducted by this website. We set five parameters, on which the results were estimated,” said Tiwari. “The prediction included the historical and current data of Indian politics. For historical, we referred to the Harvard University’s website, which has a compilation of the Lok Sabha and State Assembly elections held in India from 1977 to 2014. The current data was availed from Lokniti. Then we tried to build the model,” he said.
“For example, the regions which had a history of dominance by a particular party only, was easily predicted to have same results like the previous elections. The information gathered from the current data helped in understanding regions, where political party preferences have changed,” Tiwari told Sakal Times.
The five variables considered for the analysis are popularity of the prime ministerial candidate, the estimated vote share in different areas, the cases of vote change (the voter might have opted to vote a different party in this election from the previous one), the satisfaction level of voters from the ruling government and the priority membership (voters who follow only a particular party or have changed preferences in recent years).
Besides, the State-ruling candidates were another factor of consideration while estimating the winning probability.
Useful for political parties
Prof Akanksha Kashikar of the Statistics Department, who had guided the students in this project said, “We are not suggesting that this method is a sure-short tool to predict election results. But it can be considered as one of the options. Political parties can use it to analyse the ‘win probability’ in different areas to classify, on which section they need to focus more for gaining votes.”