Why Did Polls or OPOs Get Forecast Wrong? Perceived Bias and Inference of Bias Impact

Abstract

The failed election forecasts based on polls and Online Public Opinions(OPOs) cast doubts on the predictability of big data. In this paper, we try to examine whether polls and OPOs reveal vote preferences. Taking the 2016 US presidential election as an example, we adopt the dynamic linear models to trace two candidates' popular votes by fusing measurements from social media, news, prediction market, and polls. The interesting findings we obtained are as follows. First, with the estimations of basic dynamic linear model, we find that both the polls and OPOs are subject to measurement bias. Second, to characterize the bias, we add a time-varying parameter into the basic dynamic linear model. Via Markov Chain Monte Carlo(MCMC) estimations, we find that although polls and news failed to reflect the victory of Donald Trump, they did reflect candidates' popularity nationwide. Compared with polls and news, OPOs have larger deviations in revealing candidates' national popularity, but they are closer to the electoral vote outcome. Third, we find that there are patterns in the same type of measurement bias, no matter in direction or magnitude. The above three findings highlight the necessity of fusion method using multi-source heterogeneous data to make election forecasts. The fusion is expected to cancel out different types of measurement bias. The findings also indicate that both the polls and OPOs reveal the public opinions although the measurement bias exists. It is useful to adjust measurements extracted from polls and OPOs for specific forecast models.