Modelling German Elections with Stan by Marcus Groß (INWT)

Abstract: State-space models are a popular choice for modelling voting intentions and election results using polling data.

The presented multivariate Bayesian state-space model attempts to go beyond random-walk or Kalman-filter approaches (which only deliver comparable performance to simple weighted survey averages) by introducing a long-short term event memory effect to the problem. This effect serves as a reasonable explanation for the observation that the vote share partially tends to reverse to the party's long-term trend after larger short-term movements.

Additionally, the model accounts for government and opposition effects, house effects, poll errors and their correlations. The polling data from seven pollsters for the German national elections ("Bundestagswahl") from 1994 to 2021 is fitted with Stan and it is shown that mid- and long-term predictions of election outcomes can be considerably improved, while giving reasonable uncertainty estimates.

In a second part it is shown how to not only to model election outcomes, but also giving probability estimates on potential government coalitions by combining expert data and the polling model estimations.

Website and code: https://www.wer-gewinnt-die-wahl.de/en

Bio: Marcus Groß is a Data Scientists at INWT Statistics. He holds a PhD in Survey Statistics from the Freie Universität Berlin and loves to use Bayesian models.

The link will be published shortly before the event.