A specialized artificial intelligence model trained on proprietary data from Bridgewater Associates, the world’s largest hedge fund, has outperformed leading general-purpose AI systems like GPT, Claude, and Gemini on critical financial document tasks. This breakthrough leverages expert-labeled internal data to capture nuanced investment judgments that broader AI models cannot replicate.

Bridgewater and Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, tested their AI across six distinct financial analysis tasks that reflect routine investment work: identifying macroeconomic relevance in financial articles, detecting central bank rate change signals, segmenting boilerplate text in filings, and answering specific investor questions. The custom model achieved an average accuracy of 84.7%, surpassing the best general models’ top accuracy of 78.2% while operating at nearly fourteen times lower cost.

General AI models, although trained on massive volumes of public financial data, lack the ability to internalize the complex, tacit knowledge developed through years of analyst experience. Bridgewater’s workflows, which incorporate subtle distinctions in document relevance and investor priorities, are proprietary and not publicly available, limiting the effectiveness of generic models. For instance, in classifying financial news, the researchers refined categories beyond binary labels to include gradations like “relevant but uninteresting” versus “relevant and interesting,” mirroring the nuanced decision-making of macro investors.

The team also discovered issues with externally sourced annotations, which often contained errors. They therefore involved Bridgewater’s investment professionals to verify and correct disputed cases, embedding true expert judgment into the training set—a process impossible to achieve with prompt engineering alone. Fine-tuning with these curated labels allowed the AI to transcend mere verbal instructions and accurately replicate human investment insights.

While the results come from internal testing by Bridgewater and Thinking Machines Lab and have not undergone independent verification, they highlight a significant trend: AI systems tailored with domain-specific, expert-vetted data can outperform broadly trained frontier models in specialized tasks. This approach may redefine how AI supports investment decision-making, emphasizing the value of proprietary knowledge integration over broad data ingestion.