Do Open-Source LLMs Have a Home Bias? Investment Advice from Chinese, American, European and Thai Models (research note, replies dataset and code)

Pre-registered audit of twelve open-weight language models (Chinese, American, European, Thai and Korean) run locally with Ollama on a 240-prompt factorial design: investor nationality × prompt language × risk tolerance × horizon × template. Each model allocates 100% across eight equity markets. The archive holds the research note (paper/paper.pdf), the pre-analysis plan, the prompt set, all raw replies, the parsed panel, the estimation code and the generated tables and figures. Every number in the note is rebuilt from the cached replies with make all. Headline result: with the prompt held fixed, a model's home market receives −0.42 pp (95% CI −0.59 to −0.25) relative to what models of other origins give it, so the pre-registered home-bias hypothesis is not supported. The investor's stated country of residence moves allocations by +1.88 pp, about four times the absolute size of the home effect. Licences: code and text Apache-2.0; prompt set, parsed tables and documentation CC BY 4.0; model outputs remain subject to the licence of the model that generated them; the EXAONE-3.5 replies are research-only under the EXAONE AI Model License 1.1-NC and are not covered by CC BY 4.0.