When Bayesian Model Averaging Misleads: Overfitting in Small-N, High-P Political Forecasting

Abstract

We replicate 247 published political-event forecasts using Bayesian model averaging (BMA) and find that out-of-sample accuracy drops below random guessing when N<50 and P>25. Monte-Carlo simulations reveal that hyper-priors optimized for moderate dimensionality over-shrink signal variance, leading to 38 % false-discovery rates. A ridge-type correction restores calibration, offering a diagnostic toolkit for emerging-market forecasters.