CreditR1: Calibration-Aware Reinforcement Learning for Interpretable Corporate Credit Risk Assessment with Large Language Models

Probability-of-default (PD) estimation under the Basel and IFRS 9 frameworks requires probabilities that are both discriminative and well-calibrated. Large language models applied to credit assessment via prompting or supervised fine-tuning (SFT) yield poorly calibrated probabilities, while reinforcement learning with binary correctness rewards is structurally unsuitable for probability prediction under extreme class imbalance. We propose CreditR1, a three-stage framework: an SFT cold start on evidence-filtered reasoning chains; Group Relative Policy Optimization, guided by a composite verifiable reward combining Brier-score calibration—a strictly proper scoring rule—pairwise ranking, evidence anchoring, and format compliance; and an anti-contamination evaluation protocol. On Chinese A-share corporate credit data, CreditR1 matches gradient-boosted baselines in discrimination (AUC: 0.883±0.004 vs. 0.891 for XGBoost) while reducing expected calibration error by 24.2% versus isotonic-calibrated XGBoost (ECE: 0.047 vs. 0.062) and by 47.2% versus uncalibrated XGBoost (0.089). Because the test set contains only 119 default events, all comparisons carry firm-level bootstrap confidence intervals; the calibration advantage remains significant against every baseline after Holm–Bonferroni correction, including Platt, beta, and Bayesian-binning recalibrations. Ablations confirm each reward component is necessary. CreditR1 delivers calibrated PDs with evidence-grounded reasoning that supports internal model validation and human review; transferability beyond the Chinese A-share market remains an open empirical question.