A running reference on bias in LLM-based financial judgment: model memorization of real companies from raw financials, and how framing, naming, and batching change scores even when the underlying data is held fixed.

A reference collecting the recurring bias patterns seen when using language models to judge real financial assets, including memorization of company identity from raw figures and elicitation effects from framing, naming, and batching.

Using LLMs for Financial Applications: Bias and Considerations

Often, yes. Across seven tested models, every one correctly identified at least some real companies from raw financials with no name or sector given, which means the numbers alone can leak identity.

Can a language model identify a real company from its financial figures alone?

No. The underlying financial figures themselves can leak a company's identity even after the name is removed, so anonymization alone is not sufficient.

Is anonymizing company names enough to prevent bias in LLM financial judgments?

Yes. The same data can imply a different ordering depending on the framing, and scoring assets one at a time in isolation compresses scores into a narrower, less discriminating range than scoring them together.

Agentic Systems Cookbooks

Using LLMs for Financial Applications: Bias & Considerations