Can patterns of household purchases predict the outcome of US presidential elections?

We use NielsenIQ US retail scanner data to show that changes in sales patterns can be used to predict US presidential election results at the county level. Using a probit model, we regress 2016 election results against sales of various products six months prior to the election. We employ the results and the sales data for 2020 to forecast presidential election results in the same year. Comparison to actual election outcomes shows that our work correctly predicts election results in 86.47% of cases across 2, 602 US counties. We further study how changes in the consumption of certain goods influences voter turnout as well as Democrat and Republican votes.