Interpreting LLMs as Credit Risk Classifiers: Do Their Feature Explanations Align with Classical ML?