The traditional opinion polls are losing the capability to forecast election results because of the biasness in the sampling and the slowness in updating the polls. Although social media has a lot of real-time data, it has a considerable number of drawbacks, such as noise and demographic bias. This paper shows a new end-to-end forecasting pipeline, which is evaluated on a corpus of 1.75 million tweets related to the 2020 election in the U.S. Our approach adopts a dual-path sentiment extracted (VADER & RoBERTa) for a better accuracy and a new state level feature engineering to fix data bias. This plan transforms raw scores into 14 relative indicators, including sentiment differentials and volume ratios of tweets, which adjusts the regional activity imbalances. Average state-level prediction of a tuned Gradient Boosting Tree (GBT) classifier trained using these features was 70.6 per cent (ROC-AUC 0.69). Importantly, the cumulative prognosis was an exact duplicate of the Electoral College 306-232 majority. The feature analysis established that our relative indicators, especially the tweet volume ratio, were the strongest predictors that we engineered. The objectives in this paper are to present an alternative to traditional polling that is both powerful and easy to interpret in terms of bias reduction due to these relative characteristics. This framework has shown that it is a scalable, real-time process of political forecasting that has attained its goals of capturing the dynamics of the electoral process where conventional methods fail to do so.
Tan et al. (Sun,) studied this question.