AdaLogit: Oracle Credit Scoring via Adaptive Logistic Regression

We introduce AdaLogit, an adaptive logistic regression framework with elastic net regularization, which is particularly well suited to credit scoring applications. Unlike traditional ℓ p-regularized logistic regression, AdaLogit enjoys the oracle properties, ensuring consistent recovery of the true feature support even under complex correlation structures. Through a comprehensive empirical evaluation on synthetic and real-world credit scoring datasets, we show that AdaLogit matches or outperforms standard linear baselines while identifying significantly sparser and more stable feature sets. Compared with state-of-the-art blackbox classifiers such as TabPFN, AdaLogit approaches their predictive performance while remaining interpretable and providing better-calibrated probability estimates. For comparable balanced accuracy, AdaLogit often yields lower false negative rates than black-box alternatives. By preserving sharp tail probability calibration through oracle variable selection, this framework prioritizes default detection where error costs are highest, matching or exceeding the economic performance of TabPFN. The combination of transparency, stable feature selection and coefficient estimation, as well as accurate probability calibration, allows AdaLogit to decisively facilitate compliance with both the Basel Internal Ratings-Based (IRB) framework and the regulatory requirements of the EU Artificial Intelligence Act for high-risk AI systems.