I've been working on a project that involves measuring bias and errors created by Large Language Models in a very niche domain (the wonderful world of finance). The results are, depending on who you are, either startling or to be expected! It turns out that LLMs do indeed carry bias and preference around publicly traded stocks. This can show up in the form of a tendency to prefer [technology stocks, large-cap stocks and contrarian strategies](https://arxiv.org/html/2507.20957v4). Show a [recency bias, despite being better at gauging risk](https://arxiv.org/pdf/2409.11540). They can also show a bias towards specific markets for example [US based ChatGPT shows a systematically more optimistic view on Chinese firms than China-based DeepSeek](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5440116). On a more granular level, LLMs can express a clear bias towards certain stocks (e.g. APPL and MSFT) [despite debiasing techniques!](https://arxiv.org/html/2503.08750v1). With the difference in bias between models being incredibly diverse [across model size, generation and origin](https://arxiv.org/html/2507.20957v4). So how do you deal with this inherent bias in LLMs? Is it something you can quantify on a per model, per ticker/stock basis? Does it have an out sized impact on the market? Do stocks get a boost due to favorable bias ? Is this bias a trade-able signal?? Keep this in mind when doing a little stock picking yourself... or just buy the index (: