An AI Trained on Wall Street’s Own Data Just Beat GPT

An AI Trained on Wall Street’s Own Data Just Beat GPT

General-purpose AI models have been trained on enormous volumes of public financial data. What they cannot replicate is the judgment an experienced investment analyst builds over years — which signals matter, which documents are worth reading and how to separate the relevant from the noise inside a 300-page filing.

Bridgewater Associates, the world’s largest hedge fund by AUM, and Thinking Machines Lab, the artificial intelligence startup founded by former Open AI Chief Technology Officer Mira Murati, published research showing that encoding it directly into a model changes the outcome. Their custom model, trained on Bridgewater’s own expert-labeled data, outperformed GPT, Claude and Gemini on six financial document tasks, hitting 84.7% average accuracy against 78.2% for the best frontier model tested, at 13.8 times lower cost per task to run. Frontier models failed not because they lacked general financial knowledge, but because the correct answers depended on Bridgewater’s private workflows.

Those figures come from Bridgewater and Thinking Machines Lab’s own testing and have not been independently verified. Both companies have a commercial interest in the results, The Decoder noted.

Frontier Models Averaged 50% Accuracy on Bridgewater’s Investment Tasks

The six tasks Bridgewater defined are the routine work of investment analysis: classifying whether a financial article is relevant to a macro investor, whether a central bank document signals a rate change, where boilerplate ends in a long filing and whether a research note answers a specific investor question. Variants of GPT, Claude and Gemini averaged roughly 50% accuracy with a basic prompt, according to Thinking Machines Lab. The correct answers to these tasks depended on Bridgewater’s private investment workflows. They were never publicly available for frontier models to learn from.

The team tried the standard fix. Investment experts wrote detailed prompt instructions and reframed the tasks to match how investors think. For article classification, they replaced a binary label with three categories: relevant and interesting, relevant but uninteresting and irrelevant. That distinction captures why a small initial public offering (IPO) and a China tariff announcement carry very different weight for a macro investor. Frontier model accuracy climbed into the mid-to-high 70s and stopped at 78.2%, still below the 80% threshold Bridgewater’s investors required before trusting the system in daily work, Thinking Machines Lab reported.

A prompt can only convey what an expert can put into words. Fine-tuning, which adapts a model by training it on labeled examples rather than instructing it through prompts, transfers judgment that resists verbal description. The research team sourced initial document labels from outside contractors, found many were wrong and sent disputed cases to Bridgewater’s investment professionals for correction, Thinking Machines Lab reported. The final dataset encoded how Bridgewater investors filter information, not how the public writes about finance.

Financial Firms Are Building Domain Models on Data They Already Own

The custom model was built on Qwen3-235B, an open-source foundation model from Alibaba, using Thinking Machines Lab’s Tinker platform, which handles the infrastructure for adapting a base model to a specific task, Thinking Machines Lab reported. Murati left OpenAI as its chief technology officer in September 2024 and launched Thinking Machines Lab in February 2025. Tinker is its commercial product. The Bridgewater collaboration is its most prominent proof of concept.

Bridgewater is not the only financial firm taking this approach. Mastercard built a domain-specific model on its own transaction data that outperformed standard industry techniques, identifying infrequent but legitimate transactions that existing systems flag incorrectly, PYMNTS reported. Nvidia’s 2026 State of AI in Financial Services found 65% of financial institutions already use AI, with nearly 90% deploying or assessing it. The recurring obstacle across both is data integration rather than model capability.

Whether the accuracy edge holds as filings, central bank language and regulatory frameworks evolve is a question the research team acknowledged but did not answer. Bridgewater has described the broader program as an effort to build a complete AI investor. The six-task paper, published June 30, is the first detailed look at how it is doing that.