Forecasting Brazilian presidential elections: Solving the N problem

The use of election forecasting models is common practice in the US and other established democracies like France and the UK. However, not much work has been done in the area for more recent democracies. Forecasting election results in recently (re)democratized countries poses a serious challenge, given the very few observations of the dependent variable. Thus, we ask: is it possible to make valid election forecasts when the number of elections we have is very small? In this paper, we present recommendations on how to forecast elections under such circumstances. Our strongest recommendation is to evaluate forecasting models using subnational data. We illustrate our recommendations using Brazilian presidential elections since 1994 and data from the 27 states of the Union. Our findings indicate that forecasting elections in recent democracies is neither futile nor impossible, as some of the models presented here produce reasonably accurate forecasts.