Reasoning Externalization for Faithful Large Language Model Narratives of Stock Return Predictions

Externalizing reasoning significantly improves LLM accuracy in financial explanations: When researchers provided LLMs with structured, step-by-step reasoning about numerical changes and feature relationships rather than raw data, evidence faithfulness improved dramatically—from 0.696 to 0.996 for the Qwen3-32B-Instruct model—demonstrating that breaking down complex reasoning helps AI produce more reliable narratives.

LLMs struggle with numerical inference in finance without external support: The paper identifies a specific vulnerability where LLMs make errors when inferring numerical changes and feature relationships in stock predictions, highlighting why financial institutions cannot simply rely on LLMs to translate AI outputs without additional safeguards.

Historical context improves human trust even when automatic metrics don't capture it: While providing historical analogs (past periods with similar market conditions) didn't improve automated faithfulness scores, humans rated these explanations as significantly more useful, suggesting that trustworthiness involves factors beyond what standard verification metrics measure.

Verifiable evidence trails are essential for responsible AI narratives: The framework's requirement that "each generated claim is verified against provenance-linked evidence" demonstrates that responsible AI explanations in high-stakes domains like finance need transparent, traceable reasoning rather than black-box natural language generation.

Combining multiple explanation techniques creates more reliable AI interpretability: The study shows that temporal SHAP analysis combined with historical regime comparisons produces better outcomes than any single explanation method alone, suggesting responsible AI requires layered, complementary approaches rather than single solutions.