A New Paradigm for Supervised Machine Learning in Political Science: Tabular Foundation Models

Political scientists often employ supervised machine learning to predict outcomes from tabular data in applied research—for example, when forecasting armed conflict, election results, or judicial decisions, or when measuring variables of interest for downstream analyses. Despite advances in deep learning, gradient-boosted decision trees (GBDTs) have been the dominant predictive model for tabular data over the past decades. However, boosting algorithms require extensive hyperparameter tuning, which can be time-consuming and computationally intensive. I argue that generative tabular foundation models can rival and, in many cases, outperform existing approaches out-of-the-box without the need for hyperparameter tuning. This new class of models leverages a generative pretraining paradigm to deliver strong performance with minimal task-specific optimization or feature engineering, enabling researchers to focus on substantive questions rather than model configuration. By reanalyzing datasets from published political science studies, I show that tabular foundation models match or surpass state-of-the-art predictive accuracy while substantially reducing researcher effort and computational cost. The findings suggest that tabular foundation models can advance methodological practice in political science by lowering barriers to entry, reducing researcher degrees of freedom, and enabling more efficient predictive analyses.