March 15, 2026 · 2 min read
Calibrated Ensembles & Focal Loss for Ultra-Imbalanced Tabular Risk
Empirical study on combining PyTorch Deep Tabular ResNets with Tree Boosters and Platt probability calibration under <0.18% fraud skew.
machine-learning
imbalanced-learning
deep-learning
probability-calibration
fintech