Minimal models for forecasting U.S. elections
Election forecasting is a high-stakes problem characterized by uncertainty, incomplete information, and subjective choices (e.g., related to defining likely voters). Interestingly, polling errors tend to be correlated in states that have similiar demographics, adding further complexity. With the goal of shedding light on the forecasting process and exploring state-state interactions, we are developing data-driven minimal models for forecasting U.S. elections. Our compartmental approach tracks the number of Democrat, Republican, and undecided voters in each state, and our parameters are fit to data from HuffPost and RealClearPolitics.com.
Our model forecast the 2018 senate and governor races with the same accuracy as more complex popular sources (e.g., FiveThirtyEight.com), and we are using our model to study how subjective choices on how to account for uncertainty impact forecast interpretations.