Auditing Asset-Specific Preferences in Financial LLMs: Hidden Bitcoin Bias in Portfolio Allocation

arXiv·medium signal

Tests whether LLMs powering robo-advisors and trading agents carry built-in biases toward specific financial instruments. Finds systematic asset preferences — including measurable Bitcoin bias — that persist across prompting strategies. Examines whether internal representations encode asset-specific sentiment and whether this biases portfolio allocation. Critical for fintech builders: if your LLM-based advisor has hidden asset preferences, that's a regulatory and fiduciary risk.