Election forecasting has a long tradition, especially in American and French presidential elections (Lewis-Beck, 2005), but not much has been done elsewhere, including, most obviously, in the more recently democratized parts of the world, where elections are few and far in-between. The most important constraint in forecasting election results in the newly democratized world is the low number of observations on the dependent variable: that is, there are very few election results to build any forecast model. Thus we ask: is it possible to make valid election forecast when the number of elections is extremely low? A low N constitutes an important limitation for forecasting election results, but we believe the task is not futile nor entirely impossible. In this paper, we present recommendations on how to forecast elections when the number of elections informing our statistical models is low. We illustrate our recommendations using Brazilian presidential elections. Brazil''s recent (re)democratization in the 1980s has now produced six presidential elections (1989, 1994, 1998, 2002, 2006, and 2010). We test forecast models that attempt to explain vote choice for the incumbent candidate in the 1994 elections onward. This model, because of the low N problem, focuses on explaining election outcomes at the subnational level where we have many more observation points. More specifically, we present models that we test at the state level (Brazil has 27 states) for each of the past five elections (1994-2010).
