India Forecasting After-Action Analysis


After five weeks of voting in seven different phases, the 2019 general election in India concluded and delivered a decisive win for the NDA coalition, led by Prime Minister Narendra Modi, to form the government. In the lead up to the election, we produced the Øptimus Election Model of the Indian Lok Sabha Election. Now that the election is over, we can take a look at how the model performed. Our analysis discusses our Member of Parliament (MP) seat predictions for the two major Indian coalitions, the National Democratic Alliance (NDA, Modi-led) and the United Progressive Alliance (UPA, Gandhi-led), in the Lok Sabha (Lower House). We also included our vote-share prediction model for the Bharatiya Janata Party (BJP) by state as well as geographic regions. Our vote share predictions slightly overstated the BJP share, while the seat projections with CESS underestimated the NDA seat count. Overall, though, the model was closer to the final result than most traditional projections of the vote count, accurately predicting that BJP would do better than in 2014.


Vote Share Prediction

While this was our first attempt into forecasting the Lok Sabha election, we are overall pleased with how our predictions lined up with the final results. Compared to other publicly released opinion polls, which were widely cynical toward the BJP, forecasting their under-performance compared to 2014 results, our model predicted the BJP to slightly overperform compared to its 2014 numbers. For vote-share prediction, we utilized Multilevel Regression and Post-Stratification (MRP) to produce state-level estimates using individual-level data from a large national sample. We predicted that the BJP would gather 42.16% of the national votes cast. The BJP ended up securing 37.36% of vote share nationwide, and even exceeded 50% of vote share in 13 different states. Our model projected the BJP to cross 50% mark in 8 of those 13 states. Although our model slightly over-favored the BJP, our vote share prediction is significantly closer to the actual vote count than most other polls and forecasts. At the regional level, our predictions were within 2 percentage points in the Central and North regions.  Here is how our BJP vote share prediction stacks up against the results in six different geographic regions.


Seat Prediction

In addition to estimating vote share, we attempted to forecast the number of seats obtained by two major coalitions (NDA and UPA) and other alliances. In collaboration with the Center for Experimental Social Sciences (CESS) at University of Oxford, we produced a forecast for the number of seats each coalition would collect. Our model predicted that the ruling NDA coalition would continue to hold the majority in the Lok Sabha. To build our forecast model, our method relied on collected information from our in-house online survey, supplemental MTurk response from India, historical election results, and publicly available opinion polls. This is different from our vote share prediction where our model was solely built on survey data and combined with census data to calculate estimates. Our final forecast had NDA winning 304 seats, which is 32 more seats than the 272 majority required to form the government. We projected UPA to win 119 seats, with the remaining 120 seats divided among other parties and alliances. Here is how our seat prediction forecast stacks up against the final results.

Final Takeaways

Overall, the model caught the surge in BJP support, though the two different approaches differed on just how much strength the BJP would gain. The Indian context presents issues that are not present in forecasting American elections, which we parsed through as we were developing the modeling approaches. Unlike in the United States, voters over the age of 55 are difficult to survey in India. We employed post-stratification techniques to weight older voters at their correct level of representation, but the raw sample collected was disproportionately young and urban. The Indian Census, while very helpful in formulating demographic estimates and weights, is also much more limited in depth than the US Census we are used to employing in the American context. Despite these pitfalls, however, we are pleased with our predictions and will look to incorporate additional components in our model to address under-representation and segment gaps in the future.