A Unified Theory of Prediction Uncertainty Closes a Decades-Old Gap

This research generalizes the classical bias-variance decomposition—previously limited to squared error—to all strictly proper scoring rules, providing a unified theoretical framework for understanding model uncertainty. The theory explains why ensemble methods work and enables reliable confidence estimation even under dataset shift, solving a fundamental problem that has restricted confidence quantification to squared-error settings. This breakthrough allows practitioners to build more trustworthy predictive systems across diverse loss functions and real-world deployment scenarios.