Ranking Prior Alignment: A Model-Agnostic External Regularization Framework for Cold-Start Credit Risk Modeling

01 ABSTRACT

To address cold-start credit scoring with scarce labeled data, immature feature pipelines, and limited model capacity, the authors propose Ranking Prior Alignment, which injects external ranking priors into scoring models via a temperature-scaled KL divergence loss. On an industrial dataset of over 1.5M merchants, MIL alignment achieves 7/7 positive evaluation cells at 3K bags, XGBoost ablation achieves 9/9 positive metrics at 300 bags, and cross-dataset validation on public Amex shows 5/5 positive folds (avg Delta AUC = +0.041). The authors argue the method is model-agnostic and prior-source-independent, observe an inverse-scaling pattern where alignment gains grow as data abundance, model capacity, and feature quality decrease, and suggest this helps practitioners assess when to invest in prior annotation.

02 KEY FINDINGS

1. The framework distills external ranking priors (from domain experts, teacher models, or LLMs) into scoring models via a temperature-scaled KL divergence loss, unifying neural MIL attention and tree-based XGBoost custom objective implementations.
2. Industrial data experiments show 7/7 positive evaluation cells for MIL alignment at 3K bags (peak Delta AUC = +0.020 on OOT-1) and 9/9 positive metrics for XGBoost ablation at 300 bags; public Amex cross-validation shows 5/5 positive folds (avg Delta AUC = +0.041).
3. The authors report an inverse-scaling pattern: alignment gains grow as data abundance N, model capacity C, and feature quality Q decrease; the tree-based instantiation tolerates annotation noise up to eta = 0.5.