Triumph of the 'Quants'?: forecasting election outcomes - Simon Jackman

Simon Jackman

Abstract

Along with Nate Silver, Drew Linzer and Sam Wang, I was one of several "quants" who accurately predicted the 2012 U.S. presidential election outcome in every state. I present details of the model - a multivariate, dynamic linear model - which provides estimates of the trajectory of multiple latent targets (state-by-day levels of voting intentions) in the months leading up to the election. Among other things, the model lets us assess the impact of set piece of the campaign (conventions, debates) and exogenous events (Hurricane Sandy).

Although hailed as a “triumph of the quants”, forecasting models that relied heavily on the polls nonetheless made substantial errors in 2012. Collectively, the polling industry underestimated Obama's two-party vote share by about half a percentage point; I examine variation in polling house bias both across states and survey houses. I also assess whether pollsters “herd” in the final weeks of a campaign, with published poll estimates converging around an industry consensus, displaying less dispersion than we see in earlier stages of the campaign. Finally, I also contrast the predictive performance of poll-based forecasting models against naive, simple alternatives: the poll-based models outperform these alternative models, but not by much.