findr: Transparent and Fair Credit Risk Decisions Through Semi-Structured Regressions

Battle Count: 70

The paper is primarily focused on credit risk modeling and lending decisions rather than quantitative trading. However, the semi-structured regression framework, fairness-aware modeling, and interpretability diagnostics could be relevant for credit portfolio risk management, counterparty risk assessment, and regulatory compliance in financial institutions. The accuracy-fairness frontier methodology could inform risk model governance. Limited direct applicability to trading strategies or market prediction.

The framework requires implementing: (1) QR decomposition for orthogonalization of neural encoder outputs, (2) Wasserstein distance computation with exact quantile representation for 1D scores, (3) combined loss with fairness penalty and gradient backpropagation through both components, (4) score-level frontier construction via iterative optimization, (5) three interpretability diagnostics (EVR, DDR, CSER). The architecture is modular but requires careful handling of the orthogonalization step and Wasserstein gradient computation. Hyperparameter tuning across lambda grid adds complexity.

The paper provides detailed mathematical formulations, hyperparameter search spaces (Table 8), training procedures (AdamW, binary cross-entropy with logits, 6000 epochs, early stopping), and uses eight publicly available credit datasets. However, no code repository is explicitly mentioned. The methodology is fully specified with equations for orthogonalization, Wasserstein penalty, and frontier construction. Appendix A provides computational details for the predictor frontier.