Implementing domain-specific LLMs for strategic investment decisions: a retrospective case study comparing AI and human expertise

This study examines whether domain-specific large language models can replicate elite investment expertise through systematic comparison with Berkshire Hathaway’s performance during 2022–2024. Three LLM configurations—OpenAI GPT-4, Anthropic Claude Opus, and a consensus approach—made investment decisions at 21 predetermined points while processing identical information available to Berkshire. Although the LLM strategies demonstrated sophisticated financial training and superior risk management (maximum drawdowns of 5–7% versus 14%), they underperformed substantially, achieving cumulative returns of 4.72–18.01% compared to Berkshire’s 42.12%. Factor attribution analysis showed that LLMs generated positive security selection alpha (1.8–2.4% annually) but suffered from systematic biases including excessive momentum tilts, low market beta (0.10–0.32 versus 0.85), and premature profit-taking. Portfolio overlap of only 6–11% reflects fundamentally different investment philosophies. These findings indicate that current domain-specific LLMs lack the judgment, patience, and conviction required for superior long-term investing, supporting augmentation rather than replacement strategies in investment management.