Bridgewater, Thinking Machines Fine-Tune AI Models to Improve Financial Judgment

Bridgewater Associates’ AIA Labs partnered with AI startup Thinking Machines to test whether frontier language models could identify signals that genuinely affect investment decisions amid vast amounts of financial information. The research found that models including ChatGPT and Claude performed close to random guessing when given only general prompts, underscoring the difficulty of using general-purpose AI as a direct substitute for professional financial judgment.

The latest research showed that frontier language models achieved a baseline accuracy of only about 50% on financial-information screening tasks. After the team fine-tuned the models using a high-quality dataset labeled by professional investors, accuracy rose to 84.7%. Bridgewater said the fine-tuned models outperformed closed-source models in both accuracy and computing cost, though publicly available information did not disclose the date of the partnership or the amount invested.