🧭 What the model does

A hierarchical Dirichlet regression model with Gaussian process priors produces accurate, well-calibrated forecasts of U.S. Senate vote shares across different time horizons. The model blends time-varying opinion-poll signals with structural fundamentals and produces uncertainty estimates grounded in historical election and poll data.

📊 How polls and fundamentals are combined

• The model dynamically balances predictions coming from time-dependent opinion polls against those coming from fundamentals.
• Gaussian process priors capture temporal structure in polling trajectories, while the hierarchical Dirichlet regression maps those signals to vote-share outcomes.
• Uncertainty is derived naturally from historical patterns in elections and poll behavior, yielding credible coverage intervals for forecasts.

🧪 Tests, validation, and performance

• Experiments demonstrate high accuracy and well-calibrated coverage rates for vote-share predictions at multiple forecasting horizons.
• Validation includes a retrospective forecast of the 2018 Senate cycle and a true out-of-sample forecast for 2020.
• The approach achieves state-of-the-art accuracy and coverage despite using relatively few covariates.

💡 Why this matters

This method offers a practical, principled way to produce timely Senate forecasts that quantify uncertainty and adapt to changing polling information, making it useful for researchers and practitioners who need reliable election predictions without large sets of predictors.