Identifying and Mitigating Bias in Large Language Model-Based Financial Investment Analysis

Abstract

Large language models can assist in financial statement interpretation, risk assessment, and investment recommendation generation, but their output may be affected by corpus distribution, prompting methods, and knowledge timeliness. This paper employs a narrative review method based on structured retrieval to review research on financial scenarios and related general methods, comparing techniques such as evidence verification, counterfactual testing, uncertainty estimation, prompting bias mitigation, and Retrieval-Augmented Generation (RAG). Research shows that factual accuracy, judgment consistency, and investment effectiveness belong to different evaluation levels, and improvement in a single indicator is insufficient to prove the elimination of systematic bias. Future research should strengthen time-consistent data governance, cross-market validation, and human-machine collaborative auditing, and establish an evaluation system that combines bias identification with decision constraints.