Biased While Betting on Bernie? A cross-sectional and panel data analysis of prediction error in U.S. election prediction markets from 2015 to 2020

Abstract: A consensus seems to have emerged that political prediction markets can lose predictive power when certain efficiency criteria are not met. With a cross-sectional dataset of 570 prediction markets about U.S. elections and a panel dataset with 6,465 days of trading from PredictIt.org, I use OLS and correlated random effects models to test whether systematic prediction error is measurable under conditions of questionable market efficiency. In particular, I investigate whether candidates who share traders' ideology and elections that see high levels of voter enthusiasm are associated with higher prediction error due to wishful thinking bias. I also explore whether female and ethnic minority candidates are associated with increased prediction error due to misperceptions about their electability. Finally, I test hypotheses about how prediction error evolves over time and with changes in Google search volume. I do not find strong evidence that any of ideology, enthusiasm, gender, and ethnicity are associated with increased prediction error in these markets. The hypothesis that predictions made further away from election day see more prediction error over a short timespan is strongly supported, but I find no evidence that the duration of trading over a longer timespan or changes in Google search volume matter.