Combining survey questions with a Bayesian bootstrap method improves election forecasts

We present a Bayesian bootstrap method for election forecasting that integrates traditional own-intention questions with wisdom-of-crowds questions, which ask for social-circle expectations and election forecasts. Across the 2018, 2020, and 2024 U.S. elections, wisdom-of-crowds questions consistently outperformed own-intention questions in predicting vote shares. The Bayesian bootstrap combines these questions in a theoretically justified way, assigning weights based on their inferred informational value. The main assumption is that election winner expectations for a given participant are themselves an optimal Bayesian forecast in light of all evidence available to that participant. Social-circle expectations contribute the most to the Bayesian bootstrap forecast, as they reduce demographic biases and capture hidden voter preferences. The method improves forecast accuracy across diverse electoral contexts, offering a principled approach to combining polling questions and suggesting applicability to forecasting other social phenomena.