Communicating Credit Risk with Large Language Models: Evaluation of Explanations from Standard and Alternative Data-Based Models

Battle Count: 71

The paper focuses on credit risk decisioning and regulatory compliance rather than quantitative trading. While it involves risk management and classification models (XGBoost, GNN), the application domain is consumer lending/mortgage approval rather than trading strategies, portfolio optimization, or market prediction. The LLM-as-explanation-layer concept could tangentially inform model governance in trading systems, but the core contribution is not directly applicable to quantitative trading workflows.

The pipeline involves multiple complex components: XGBoost training with hyperparameter tuning, GNN (GAT) construction with network building from loan data, SHAP and GNNExplainer post-hoc explanations, LoRA fine-tuning of LLMs (Gemma 3 4B and DeepSeek R1 70B), structured prompt engineering with multiple strategies, LLM-as-a-judge evaluation, mixed-effects statistical modeling, and linguistic feature analysis. The bimodal pipeline requires careful evidence integration. However, the modular architecture and available code repository reduce practical barriers.

Code and survey data are available on GitHub (https://github.com/Banking-Analytics-Lab/LLMExplainer). The dataset (Freddie Mac single-family loan-level data) is publicly available. Prompt templates, fine-tuning configurations, and evaluation procedures are described in detail. However, some LLM configurations (Gemini 2.5 API) may have version drift, and the synthetic supervision dataset generation via ChatGPT-4o introduces some non-determinism. Statistical methods are fully specified in appendices.