Calibrated Credit Intelligence: Shift-Robust and Fair Risk Scoring with Bayesian Uncertainty and Gradient Boosting

Battle Count: 71

The paper is primarily focused on credit risk scoring rather than trading strategies. However, its contributions in distribution shift robustness, calibrated probability estimation, uncertainty quantification, and fairness-aware modeling are transferable to quantitative trading contexts (e.g., credit-linked instruments, lending portfolio management, risk-adjusted signal generation). The emphasis on model stability under temporal drift and reliable probability outputs is relevant for any financial ML deployment.

The CCI pipeline involves multiple complex components: (1) training a Bayesian Neural Network with variational inference (ELBO optimization), (2) training a fairness-constrained GBDT with a custom regularized objective, (3) implementing a drift detection mechanism and adaptive fusion weight selection, (4) post-hoc temperature scaling calibration, (5) fairness auditing across sensitive groups, and (6) SHAP-based explainability. The multi-stage pipeline with interdependent components, time-consistent data splitting, and the need for careful hyperparameter tuning across both BNN and GBDT components makes implementation non-trivial. No reference code is provided.

The paper uses a public Kaggle dataset (Home Credit Credit Risk Model Stability), provides detailed algorithm pseudocode (Algorithm 1), explicit mathematical formulations for all components, and specifies preprocessing steps. However, no GitHub repository or code link is provided, and hyperparameter details (e.g., exact BNN architecture, number of MC samples S, fairness tolerance Δ_max values) are not fully specified. Results are reported with mean±std over multiple runs, which aids reproducibility.